CN110352297B - machine learning device - Google Patents

Abstract

Appropriate output values can be obtained even when the values of the operating parameters are outside the preset range. In a machine learning device for using a neural network to output an output value with respect to a value of an operating parameter of a device, when the value of the operating parameter of the device is outside a preset range, the front end of the output layer of the neural network is increased. The number of nodes in one hidden layer is based on the training data obtained by actual measurement of the newly acquired values of the operating parameters of the machine, so that the output values that change according to the values of the operating parameters of the machine are the same as the values corresponding to the operating parameters of the machine. The weights of the neural network are learned in such a way that the difference between the values of the training data becomes smaller.

Description

machine learning device

technical field

The present invention relates to machine learning devices.

Background technique

Among the control devices for internal combustion engines using neural networks, there are known control devices for internal combustion engines that adjust the amount of intake gas into the combustion chamber to the actual amount based on values of operating parameters of the internal combustion engine, such as the engine speed and the amount of intake air. The weights of the neural network are learned in advance so that the amount of intake gas in the combustion chamber is the same, and the amount of intake gas into the combustion chamber is estimated from the values of the operating parameters of the internal combustion engine using the neural network that has learned the weights during the operation of the internal combustion engine (for example, refer to Patent Documents). 1).

prior art literature

Patent Literature

Patent Document 1: Japanese Patent Laid-Open No. 2012-112277

SUMMARY OF THE INVENTION

The problem to be solved by the invention

In addition, the use range of the value of a specific type of operating parameter related to the internal combustion engine, such as the engine speed, can be pre-estimated according to the type of the internal combustion engine. Therefore, the pre-estimated use range of the value of the operating parameter of the internal combustion engine is usually used so that the nerve The weights of the neural network are learned in advance so that the difference between the output value of the network and the actual value such as the actual intake gas amount into the combustion chamber becomes small. However, in practice, the values of the operating parameters of the internal combustion engine may be outside the pre-estimated range of use. In this case, since learning based on the actual values is not performed for the out-of-premise use range, there is a possibility of using a neural network calculation. There is a problem that the output value will be a value that deviates greatly from the actual value. Such a problem is not limited to the field of internal combustion engines, but occurs in machines of various fields that are targets of machine learning.

In order to solve the above-mentioned problem, according to the first invention, there is provided a machine learning apparatus for outputting an output value with respect to a value of an operation parameter of a machine using a neural network, wherein a specific parameter related to the above-mentioned machine is preset. The value range of the operating parameter of the category, and the number of nodes of the hidden layer of the neural network corresponding to the value range of the operating parameter of the specific category related to the above-mentioned machine is preset. When the value of the operating parameter of the specific type related to the machine is outside the preset range, the number of nodes in the previous hidden layer of the output layer of the neural network is increased, and the newly acquired specific type related to the machine is used. The training data obtained by the actual measurement of the values of the operating parameters of the category and the training data obtained by the actual measurement of the values of the operating parameters of the above-mentioned equipment within a preset range are used to learn the weights of the neural network, and the neural network that has learned the weights is used. The network outputs output values relative to the values of the specific classes of operating parameters associated with the above-mentioned machines.

In order to solve the above-mentioned problems, according to the second invention, there is provided a machine learning apparatus for outputting an output value with respect to a value of an operating parameter of a machine using a neural network, wherein a plurality of parameters related to the above-mentioned machine are preset. The range of the values of the operating parameters of each category, and the number of nodes in the hidden layer of the neural network corresponding to the value ranges of the operating parameters of the plurality of categories related to the above-mentioned equipment is preset, and the newly acquired and When the values of the operating parameters of the above-mentioned equipment-related categories are outside the preset range, increase the number of nodes in the previous hidden layer of the output layer of the neural network, and use the newly acquired equipment related to the above-mentioned equipment. The weights of the neural network are learned using the training data obtained by the actual measurement of the values of the operating parameters of the related multiple categories and the training data obtained by the actual measurement of the values of the operating parameters of the above-mentioned equipment within a preset range. A weighted neural network is used to output an output value with respect to the values of the operating parameters of the plurality of categories related to the above-mentioned equipment.

In order to solve the above-mentioned problems, according to a third invention, there is provided a machine learning apparatus for outputting an output value with respect to a value of an operation parameter of a machine using a neural network, wherein a plurality of parameters related to the above-mentioned machine are preset. The range of values of the operating parameters of each category, and a neural network corresponding to the value range of the operating parameters of the plurality of categories related to the above-mentioned equipment is pre-formed, and the newly acquired plural categories of the above-mentioned equipment related to the above-mentioned equipment. When the value of the operating parameter of at least one category of the values of the operating parameters is outside the preset range, a new neural network is formed, and the newly acquired values of the operating parameters of the plurality of categories related to the above-mentioned equipment are used to pass the actual measurement. On the other hand, the weights of the new neural network are learned from the obtained training data, and output values corresponding to the values of the operation parameters of the plurality of categories related to the above-mentioned equipment are output using the neural network that has learned the weights.

Invention effect

In each of the above-described inventions, when the value of the newly acquired operating parameter of the device is outside the preset range, by increasing the number of nodes in the hidden layer of the neural network or by creating a new neural network, it is possible to suppress the When the value of the operating parameter of the device is out of the preset range, the output value calculated using the neural network is a value that greatly deviates from the actual value.

Description of drawings

FIG. 1 is an overall view of an internal combustion engine.

FIG. 2 is a diagram showing an example of a neural network.

3A and 3B are diagrams showing changes in the value of the sigmoid function σ.

4A and 4B are diagrams showing output values from a neural network and nodes in a hidden layer, respectively.

5A and 5B are diagrams showing output values from nodes in the hidden layer and output values from nodes in the output layer, respectively.

6A and 6B are diagrams showing the neural network and the output values from the nodes of the output layer, respectively.

7A and 7B are diagrams for explaining the problem to be solved by the invention of the present application.

8A and 8B are diagrams showing a neural network and a relationship between input values and output values of the neural network, respectively.

FIG. 9 is a diagram illustrating a neural network.

FIG. 10 is a flowchart for performing learning processing.

FIG. 11 is a diagram showing a modification of the neural network.

FIG. 12 is a flowchart illustrating another embodiment for performing a learning process.

FIG. 13 is a diagram showing a neural network.

14A and 14B are diagrams showing preset ranges of the engine speed and the like.

FIG. 15 is a diagram showing a modification of the neural network.

FIG. 16 is a flowchart showing yet another embodiment for performing the learning process.

FIG. 17 is a diagram showing learned divided regions divided according to the values of the operating parameters of the internal combustion engine.

18A , 18B and 18C respectively show the distribution of training data with respect to the engine speed and ignition timing, the distribution of training data with respect to ignition timing and throttle opening, and the relationship between the training data and the output value after learning 's diagram.

19A and 19B are diagrams showing the relationship between training data and learned output values.

FIG. 20 is an overall view of a machine learning device for performing automatic adjustment of an air conditioner.

FIG. 21 is a diagram showing a neural network.

22A and 22B are diagrams showing preset ranges of air temperature and the like.

FIG. 23 is a flowchart showing yet another embodiment for performing the learning process.

24A and 24B are diagrams showing preset ranges of air temperature and the like.

FIG. 25 is a flowchart showing yet another embodiment for performing a learning process.

FIG. 26 is an overall view of a machine learning apparatus for estimating the degree of deterioration of a secondary battery.

FIG. 27 is a diagram illustrating a neural network.

28A and 28B are diagrams showing preset ranges of air temperature and the like.

FIG. 29 is a flowchart for executing calculation processing.

FIG. 30 is a flowchart for executing training data acquisition processing.

FIG. 31 is a flowchart showing yet another embodiment for performing a learning process.

32A and 32B are diagrams showing preset ranges of air temperature and the like.

FIG. 33 is a flowchart showing yet another embodiment for performing the learning process.

Detailed ways

<Overall structure of internal combustion engine>

First, a case where the machine learning device of the present invention is applied to an internal combustion engine will be described. Referring to FIG. 1 showing an overall view of an internal combustion engine, 1 denotes an internal combustion engine main body, 2 denotes a combustion chamber of each cylinder, 3 denotes a spark plug arranged in a combustion chamber 2 of each cylinder, and 4 denotes a fuel for supplying fuel to each cylinder (for example, Gasoline) fuel injection valve, 5 represents the balance tank, 6 represents the intake manifold, and 7 represents the exhaust manifold. The balance tank 5 is connected to the outlet of the compressor 9 a of the exhaust turbocharger 9 via the intake duct 8 , and the inlet of the compressor 9 a is connected to the air cleaner 11 via the intake air amount detector 10 . A throttle valve 12 driven by an actuator 13 is arranged in the intake duct 8 , and a throttle valve opening degree sensor 14 for detecting the throttle valve opening degree is attached to the throttle valve 12 . In addition, an intercooler 15 for cooling intake air flowing in the intake duct 8 is arranged around the intake duct 8 .

On the other hand, the exhaust manifold 7 is connected to the inlet of the exhaust turbine 9 b of the exhaust turbocharger 9 , and the outlet of the exhaust turbine 9 b is connected to the exhaust gas purification catalytic converter 17 via the exhaust pipe 16 . The exhaust manifold 7 and the balance tank 5 are connected to each other via an exhaust gas recirculation (hereinafter, referred to as EGR) passage 18 , and an EGR control valve 19 is arranged in the EGR passage 18 . Each of the fuel injection valves 4 is connected to a fuel distribution pipe 20 which is connected to a fuel tank 22 via a fuel pump 21 . An _NOx sensor 23 for detecting the _NOx concentration in the exhaust gas is arranged in the exhaust pipe 16 . In addition, an atmospheric temperature sensor 24 for detecting the atmospheric temperature is arranged in the air cleaner 11 .

The electronic control unit 30 is constituted by a digital computer, and includes a ROM (Read Only Memory) 32, a RAM (Random Access Memory) 33, a CPU (Microprocessor) 34, an input port 35, and an output port, which are connected to each other via a bidirectional bus 31. 36. The output signals of the intake air amount detector 10 , the throttle opening sensor 14 , the _NOx sensor 23 , and the atmospheric temperature sensor 24 are input to the input port 35 via the corresponding AD converters 37 . A load sensor 41 that generates an output voltage proportional to the amount of depression of the accelerator pedal 40 is connected to the accelerator pedal 40 , and the output voltage of the load sensor 41 is input to the input port 35 via the corresponding AD converter 37 . Furthermore, a crank angle sensor 42 that generates an output pulse every time the crankshaft rotates, for example, 30°, is connected to the input port 35 . Within the CPU 34 , the engine speed is calculated based on the output signal of the crank angle sensor 42 . On the other hand, the output port 36 is connected to the spark plug 3 , the fuel injection valve 4 , the throttle valve driving actuator 13 , the EGR control valve 19 , and the fuel pump 21 via the corresponding drive circuit 38 .

<Outline of Neural Network>

In the embodiment of the present invention, various values representing the performance of the internal combustion engine are estimated using a neural network. FIG. 2 shows an example of this neural network. The circular symbols in Figure 2 represent artificial neurons, which are commonly referred to as nodes or units in neural networks (referred to in this application as nodes). In FIG. 2, L=1 represents the input layer, L=2 and L=3 represent the hidden layer, and L=4 represents the output layer. In addition, in FIG. 2, x ₁ and x ₂ represent the output values from the nodes of the input layer (L=1), y represents the output values from the nodes of the output layer (L=4), z ₁ , z ₂ and z ₃ represents the output value from each node of the hidden layer (L=2), and z ₁ and z ₂ represent the output value from each node of the hidden layer (L=3). It should be noted that the number of hidden layers can be set to one or any number, and the number of nodes of the input layer and the number of nodes of the hidden layer can also be set to any number. It should be noted that although FIG. 2 shows a case where the number of nodes of the output layer is one, the number of nodes of the output layer may be more than two.

In each node of the input layer, the input is output as it is. On the other hand, the output values x ₁ and x ₂ of each node of the input layer are input to each node of the hidden layer (L=2), and the corresponding weights w and 2 are used in each node of the hidden layer (L=2) Bias b to calculate the total input value u. For example, the total input value uk calculated in the node indicated by z _k ( _k =1, 2, 3) of the hidden layer (L=2) in FIG. 2 is as follows.

Next, the total input value _uk is transformed by the activation function f, and is output from the node indicated by z _k of the hidden layer (L=2) as the output value z _k (=f(u _k )). The same is true for other nodes in the hidden layer (L=2). On the other hand, the output values z 1 , z 2 and z 3 of each node of the hidden layer (L=2) are input to each node of the hidden layer (L=3), and the output values z ₁ , z ₂ and z ₃ of each node of the hidden layer (L=3) are , the total input value u(Σz·w+b) is calculated using the corresponding weight w and bias b. The total input value u is also transformed by the activation function, and is output from each node of the hidden layer (L=3) as output values z ₁ and z ₂ . It should be noted that, in the embodiment of the present invention, Sigmoid is used. The function σ serves as the activation function.

On the other hand, the output values z ₁ and z ₂ of each node of the hidden layer (L=3) are input to the nodes of the output layer (L=4), and the corresponding weights w and bias b are used in the nodes of the output layer. to calculate the total input value u(Σz·w+b), or use only the corresponding weight w to calculate the total input value u(Σz·w). In the embodiment of the present invention, the identity function is used in the nodes of the output layer, so the total input value u calculated in the nodes of the output layer is directly output as the output value y from the nodes of the output layer.

<Performance of neural network-based functions>

When a neural network is used, an arbitrary function can be expressed, and this will be briefly described next. First, the sigmoid function σ used as the activation function will be described. The sigmoid function σ is represented by σ(x)=1/(1+exp(-x)), and as shown in FIG. 3A , it takes 0 according to the value of x value between 1 and 1. Here, when x is replaced by wx+b, the sigmoid function σ is represented by σ(wx+b)=1/(1+exp(-wx-b)). Here, when the value of w is increased, as shown by the curves σ ₁ , σ ₂ , and σ ₃ in FIG. 3B , the slope of the curve portion of the sigmoid function σ(wx+b) gradually becomes steeper. is infinite, then as shown by the curve σ ₄ in FIG. 3B, the Sigmoid function σ(wx+b) becomes x=-b/w (x which becomes wx+b=0, that is, becomes σ(wx+b) = 0.5 x, as shown in Fig. 3B, changes in a step-like manner. By utilizing the properties of such a sigmoid function σ, an arbitrary function can be expressed using a neural network.

For example, as shown in FIG. 4A, an input layer (L=1) consisting of one node, a hidden layer (L=2) consisting of two nodes, and an output layer (L=3) consisting of one node are used The constructed neural network can express a function that approximates a quadratic function. It should be noted that in this case, even if the number of output layers (L=3) is set to a large number, any function can be expressed, but for ease of understanding, the number of nodes of the output layer (L=3) is used. Take one case as an example. In the neural network shown in FIG. 4A, as shown in FIG. 4A, the input value x is input to the node of the input layer (L=1), and the weight is used to input the node input shown by z ₁ in the hidden layer (L=2). The input value u=x·w ₁ ^(L2) +b ₁ calculated by w ₁ ^(L2) and the bias b ₁ . The input value u is transformed by the sigmoid function σ(x·w ₁ ^(L2) +b ₁ ), and is output as the output value z ₁ . Similarly, the input value u=x·w ₂ ^(L2) +b ₂ calculated using the weight w ₂ ^(L2) and the bias b ₂ is input to the node indicated by z ₂ in the hidden layer (L=2). u is transformed by the Sigmoid function σ(x·w ₂ ^(L2) +b ₂ ), and output as an output value z ₂ .

On the other hand, the output values z ₁ and z ₂ of each node of the hidden layer (L=2) are input to the nodes of the output layer (L=3), and the corresponding weights w ₁ ^(y ) are used for the nodes of the output layer. ⁾ and w ₂ ^(y) to calculate the total input value u (Σz·w=z ₁ ·w ₁ ^(y) +z ₂ ·w ₂ ^(y) ). As mentioned above, in the embodiment of the present invention, the identity function is used in the nodes of the output layer. Therefore, the total input value u calculated in the nodes of the output layer is directly used as the output value y from the nodes of the output layer. output.

(I) of FIG. 4B shows that the weight w ₁ ^(L2) and the bias b ₁ are set so that the value of the Sigmoid function σ(x·w ₁ ^(L2) + b ₁ ) becomes substantially zero at x=0 The output value z ₁ from the node of the hidden layer (L=2) at . On the other hand, in the Sigmoid function σ(x·w ₂ ^(L2) +b ₂ ), if, for example, the weight w ₂ ^(L2) has a negative value, the Sigmoid function σ(x·w ₂ ^(L2) +b ₂ ) The shape of the curve is a shape that decreases as x increases, as shown in (II) of FIG. 4B . (II) of FIG. 4B shows that the weight w ₂ ^(L2) and the bias b ₂ are set so that the value of the Sigmoid function σ(x·w ₂ ^(L2) + b ₂ ) becomes substantially zero at x=0 The change in the output value z ₂ of the node from the hidden layer (L=2) at .

On the other hand, in (III) of FIG. 4B , the sum (z ₁ +z ₂ ) of output values z ₁ and z ₂ from each node of the hidden layer (L=2) is shown by a solid line. It should be noted that, as shown in FIG. 4A , the respective output values z ₁ and z ₂ are multiplied by their corresponding weights w ₁ ^(y) and w ₂ ^(y) , and in (III) of FIG. 4B , the dotted line A is used. Changes in the output value y when w ₁ ^(y) , w ₂ ^(y) >1 and w ₁ ^(y) ≈ w ₂ ^(y) are shown. Furthermore, in (III) of FIG. 4B , the output value y when w ₁ ^(y) , w ₂ ^(y) >1, and w ₁ ^(y) >w ₂ ^(y) is shown by a one-dot chain line B The change of , in (III) of FIG. 4B , the output values when w ₁ ^(y) , w ₂ ^(y) >1 and w ₁ ^(y) <w ₂ ^(y) are shown with a dashed-dotted line C change in y. In (III) of FIG. 4B , the shape of the broken line A in the range shown by W represents a curve approximated to a quadratic function as shown by y=ax ² (a is a coefficient), so it can be seen that by using FIG. 4A The neural network shown is capable of expressing a function that approximates a quadratic function.

On the other hand, FIG. 5A shows a case where the value of the sigmoid function σ is changed stepwise as shown in FIG. 3B by increasing the values of the weights w ₁ ^(L2) and w ₂ ^(L2) in FIG. 4A . (I) of FIG. 5A shows that the value of the Sigmoid function σ(x·w 1 (L2) +b 1 ) at x=−b ₁ /w ₁ ^(L2) is set so that the value of the sigmoid function σ(x·w ₁ ^(L2) +b ₁ ) increases in a step-like manner The output value z ₁ of the node from the hidden layer (L=2) at the weight w ₁ ^(L2) and the bias b ₁ . In addition, (II) of FIG. 5A shows the Sigmoid function σ(x·w ₂ ^{( L2)} at x=−b ₂ /w ₂ (L2) slightly larger than x=−b ₁ /w ₁ ^{(L2) )} +b ₂ ) is set so that the value of the weight w ₂ ^(L2) and the output value z ₂ from the node of the hidden layer (L=2) when the bias b ₂ is reduced in a stepwise manner. In addition, in (III) of FIG. 5A , the sum (z ₁ +z ₂ ) of output values z ₁ and z ₂ from each node of the hidden layer (L=2) is shown by a solid line. As shown in FIG. 4A , the respective output values z ₁ and z ₂ are multiplied by their corresponding weights w ₁ ^(y) and w ₂ ^(y) , and in (III) of FIG. 5A , w ₁ ^{( y)} , the output value y when w ₂ ^(y) >1.

In this way, in the neural network shown in FIG. 4A , a long output value y as shown in (III) of FIG. 5A is obtained through a pair of nodes in the hidden layer (L=2). Therefore, if the number of pairs of nodes in the hidden layer (L=2) is increased, and the values of the weight w and the bias b at each node of the hidden layer (L=2) are appropriately set, the expression as shown in FIG. 5B can be obtained. The approximate function y=f(x) shown by the dotted curve. In addition, in FIG. 5B, although each strip|line is drawn so that it may contact|connect, in actuality, each long strip|line may partially overlap. In addition, since the value of w is not actually infinite, each strip does not have an exact strip shape, but has a curved shape like the upper half of the curved portion shown by σ ₃ in FIG. 3B . It should be noted that although the detailed description is omitted, as shown in FIG. 6A , if a pair of nodes corresponding to two different input values x ₁ and x ₂ are set in the hidden layer (L=2), the following As shown in FIG. 6B , a columnar output value y corresponding to the input values _x1 and _x2 is obtained. In this case, it can be seen that if many pairs of nodes are provided in the hidden layer (L=2) for each of the input values x _I and x ₂ , a plurality of columns corresponding to different input values x _I and x ₂ are obtained, respectively. Therefore, a function representing the relationship between the input values x ₁ and x ₂ and the output value y can be expressed. It should be noted that even when there are three or more different input values x, a function representing the relationship between the input value x and the output value y can be expressed similarly.

<Learning in Neural Networks>

若改写则成为下式所示那样。On the other hand, in the embodiment of the present invention, the error back propagation method is used to learn the value of each weight w and the value of the bias b in the neural network. This error backpropagation method is well known, and therefore, the outline of the error backpropagation method will be briefly described below. It should be noted that the bias b is one of the weights w, and therefore, in the following description, the bias b is set to be one of the weights w. In the neural network shown in Fig. 2, when the weight in the input value u ^(L) to the node of each layer of L=2, L=3 or L=4 is represented by w ^(L) , the error function The differential of E to the weight w ^(L) is the gradient

When rewritten, it becomes as shown in the following formula.

因此若设为

则上述(1)式能够用下式表示。here,

So if set to

Then, the above-mentioned formula (1) can be represented by the following formula.

Here, when u ^(L) fluctuates, the error function E fluctuates due to changes in the total input value u ^(L+1) of the next layer, so δ ^(L) can be expressed by the following equation.

Here, when expressed as z ^(L) =f(u ^(L) ), the input value _uk ^(L+1) appearing on the right side of the above-mentioned formula (3) can be expressed by the following formula.

是δ^(L+1)，上述(3)式的右边第二项

能够用下式表示。Here, the first term on the right side of the above formula (3)

is δ ^(L+1) , the second term on the right-hand side of equation (3) above

It can be represented by the following formula.

Therefore, δ ^(L) is represented by the following formula.

which is,

That is, after δ ^(L+1) is obtained, δ ^(L) can be obtained.

When the training data y _t is obtained for a certain input value, and the output value from the output layer is y with respect to the input value, when the squared error is used as the error function, the squared error E is E=1/ 2(yy _t ) ² is found. In this case, in the node of the output layer (L=4) of FIG. 2, the output value y=f(u ^(L) ), therefore, in this case, at the node of the output layer (L=4) The value of δ ^(L) is as follows.

In the embodiment of the present invention, as described above, f(u ^(L) ) is an identity function, and f'(u ^(L) )=1. Therefore, δ ^(L) = yy _t , and δ ^{(L) is} obtained.

当求出梯度

后，使用该梯度

以使误差函数E的值减小的方式更新权重w的值。即，进行权重w的值的学习。需要说明的是，在输出层(L＝4)具有多个节点的情况下，若将来自各节点的输出值设为y₁、y₂…，将对应的训练数据设为y_t1、y_t2…，则作为误差函数E，使用以下的平方和误差E。After δ ^{(L) is} obtained, δ ^{(L-1) of the front layer is obtained using the above formula (6)} . In this way, the δ of the previous layer is successively obtained, and using the values of these δ, according to the above formula (2), the differential of the error function E, that is, the gradient, is obtained with respect to each weight w

When finding the gradient

After that, use the gradient

The value of the weight w is updated in such a way that the value of the error function E decreases. That is, the learning of the value of the weight w is performed. It should be noted that when the output layer (L=4) has a plurality of nodes, if the output value from each node is set as y ₁ , y ₂ . . . , and the corresponding training data is set as y _t1 , y _t2 ..., the following square sum error E is used as the error function E.

<Example of the present invention>

Next, a first embodiment of the machine learning apparatus of the present invention will be described with reference to FIGS. 7A to 10 . In the first embodiment of the present invention, as shown in FIG. 4A, a neural network comprising an input layer (L=1), a hidden layer (L=2) composed of one layer, and an output layer (L=3) is used network. In addition, this first embodiment shows the case where the learning of the weight of the neural network is performed so that the output value y is represented by the quadratic function of the input value x using the neural network shown in FIG. 4A . It should be noted that, in FIG. 7A to FIG. 8B , the dotted line represents the waveform of the real quadratic function, the black circle represents the training data, and the annular circle represents the output value y corresponding to the input value x and the training data. The output value y after the learning of the weight of the neural network is performed so that the difference becomes smaller, and the solid line curve represents the relationship between the input value x and the output value y after the learning is completed. In addition, in FIGS. 7A to 8B , between A and B, that is, R represents a preset range of the input value x.

FIGS. 7A and 7B are diagrams for explaining the problem to be solved by the present invention. Therefore, first, the problem to be solved by the present invention will be described with reference to FIGS. 7A and 7B . Fig. 7A shows the following situation: using a neural network in which the number of nodes in the hidden layer (L=2) is 2 as shown in Fig. 4A, for an input value x within a preset range R, the output The weights of the neural network are learned in such a way that the quantity y becomes a quadratic function y=ax ² (a is a constant) of the input value x. As shown in FIG. 7A , even when the hidden layer (L=2) of the neural network has only two nodes, when the input value x is within the range R set in advance, as shown by the solid line, the expression A function close to a quadratic function.

That is, when learning is performed with respect to the preset range R of the input value x, with respect to the preset range R, the output value y appears close to A function of a quadratic function. However, since learning is not performed outside the preset range R of the input value x, the straight line portions at both ends of the curve portion where the sigmoid function σ greatly changes appears as the output value y as shown by the solid line. Therefore, the output value y after the learning is completed, as shown by the solid line in FIG. 7A , appears in the form of a function close to the quadratic function within the preset range R of the input value x. Outside the set range R, it appears in the form of a nearly straight line that is almost unchanged with respect to the input value x. Therefore, as shown in FIG. 7A , outside the preset range R of the input value x, the output value y deviates significantly from the quadratic curve shown by the dotted line.

On the other hand, FIG. 7B shows a case where the input value x is set to x when the input value x is outside the preset range R of the input value x as indicated by x ₀ in FIG. 7B , for example. The output value y ₀ at ₀ is also included in the training data, and the weights of the neural network are learned. In this way, when learning is performed including the output value y ₀ outside the preset range R of the input value x, the straight line portion of the Sigmoid function σ shown in z ₁ in FIG. 4B that becomes z ₁ =1 is The way including the output value y ₀ rises, the Sigmoid function σ shown in z ₂ in FIG. 4B moves to the right as a whole and the value of the Sigmoid function σ becomes lower as a whole, so as shown by the solid line in FIG. In the range R, the value of the output value y after the learning is completed greatly deviates from the quadratic curve. In this way, when the input value x is outside the preset range R, an appropriate output value y cannot be obtained.

However, it was found that in this case, if the number of nodes in the hidden layer (L=2) of the neural network is increased, even if the input value x is outside the preset range R, the A suitable output value y can be obtained. Next, this will be described with reference to FIGS. 8A and 8B showing the first embodiment of the present invention. FIG. 8B shows the output value when the input value x is x ₀ in a state where the number of nodes in the hidden layer (L=2) of the neural network is increased from 2 to 3 as shown in FIG. 8A y ₀ is also included in the training data and the learning result when the weight of the neural network is learned. When the number of nodes of the hidden layer (L=2) of the neural network is increased in this way, as shown by the solid line in FIG. 8B , the value of the output value y overlaps with the quadratic curve shown by the dotted line. Therefore, as shown in FIG. 8B , it can be seen that even when the input value x is outside the pre-estimated use range R, by increasing the number of nodes in the hidden layer (L=2) of the neural network, it is possible to obtain Appropriate output value y. Therefore, in the first embodiment of the present invention, when the input value x is outside the preset range R, the number of nodes of the hidden layer (L=2) of the neural network is increased.

Next, a specific example of the input value x and the output value y shown in FIGS. 7A to 8B will be described. In the field of internal combustion engines, when the value of a specific type of operating parameter related to the internal combustion engine is the input value x, the actual output y may be in the form of a quadratic function of the input value x. As an example of such a case, There are cases where the input value x, which is a value of a specific type of operating parameter related to the internal combustion engine, is the engine speed N (rpm) and the output amount y is the exhaust gas loss amount. In this case, the use range of the engine speed N is determined correspondingly when the internal combustion engine is determined, and therefore, the range of the engine speed N is set in advance. On the other hand, the exhaust gas loss amount represents the thermal energy discharged from the combustion chamber of the internal combustion engine, and is proportional to the amount of exhaust gas discharged from the combustion chamber of the internal combustion engine and proportional to the temperature difference between the temperature of the exhaust gas discharged from the combustion chamber of the internal combustion engine and the outside air temperature. The exhaust gas loss amount is calculated based on the detected value of the gas temperature and the like when the internal combustion engine is actually operated. Therefore, the calculated exhaust gas loss amount represents a value obtained by actual measurement.

In this specific example, when the input value x, that is, the engine speed N, is within the predetermined range R, training data obtained by actual measurement is used so that the difference between the output value y and the training data corresponding to the input value x is obtained. Learning the weights of a neural network in a smaller way. That is, when the value of the operating parameter of the specific type related to the internal combustion engine is within the preset range R, the training data obtained by actual measurement is used so that the output value y corresponds to the operating parameter of the specific type related to the internal combustion engine. The weights of the neural network are learned in such a way that the difference between the values of the training data becomes smaller. On the other hand, when the input value x, that is, the engine speed N, is outside the preset range, the number of nodes in the hidden layer of the neural network is increased, and the newly acquired input value x, that is, the engine speed N, is determined by actual measurement. From the obtained training data, the weights of the neural network are learned in such a way that the difference between the output value y and the training data corresponding to the input value x becomes small. That is, when the value of the operation parameter of the specific type related to the internal combustion engine is outside the preset range, the number of nodes in the hidden layer of the neural network is increased, and the newly acquired operation of the specific type related to the internal combustion engine is used. The weights of the neural network are learned so that the difference between the output value y and the training data corresponding to the value of the specific type of operating parameter related to the internal combustion engine becomes small from the training data obtained by the actual measurement of the parameter value. Therefore, in this case, even if the engine speed N becomes higher than the preset range R, the exhaust gas loss amount can be estimated relatively accurately.

It should be noted that the first embodiment of the present invention can also be applied to a neural network having multiple hidden layers (L=2 and L=3) as shown in FIG. 9 . In the neural network shown in Fig. 9, according to the output values z ₁ , z ₂ of the nodes of the previous hidden layer (L=3) of the output layer (L=4), the output from the output layer (L=4) The form of the function is determined. That is, what kind of function expression the output value y can represent is governed by the number of nodes in the hidden layer (L=3) preceding the output layer (L=4). Therefore, in the neural network shown in Figure 9, when the number of nodes in the hidden layer is increased, as shown in Figure 9, the previous hidden layer (L=3) of the output layer (L=4) is increased. the number of nodes.

In the first embodiment described above, the amount of exhaust gas loss actually measured for various input values x within the preset range R is obtained as training data in advance, that is, the sum of the values within the preset range R and the The values of the specific types of operating parameters related to the internal combustion engine are obtained in advance by actual measurement to obtain training data, and based on the values of the specific types of operating parameters related to the internal combustion engine and the training data, the structure of the neural network is determined so that the output value y corresponds to The weights of the neural network are learned in advance so that the difference between the training data of the values of the specific types of operating parameters related to the internal combustion engine becomes small. The training data obtained in advance by actual measurement of the values of the specific types of operating parameters related to the internal combustion engine within the preset range R are stored in the storage unit of the electronic control unit 30 . In this first embodiment, a neural network having the same structure as the neural network used in the prior learning is used, and the learning is further carried out on-board while the vehicle is running, using the weights of the neural network at the time of completion of the learning. FIG. 10 shows the learning processing routine of the first embodiment in the vehicle-mounted manner. It should be noted that the learning processing routine shown in FIG. 10 is executed by interruption at regular intervals (for example, every one second).

Referring to FIG. 10 , first, in step 101 , the learned weights stored in the storage unit of the electronic control unit 30 and the training data used in the previous learning, that is, the internal combustion engine-related weights within the preset range R are read. The values A and B representing the range R of the input data, that is, the predetermined range of the values of the specific types of operating parameters related to the internal combustion engine, are training data obtained in advance by actual measurement. This learned weight is used as an initial value for the weight. Next, in step 102, the number K of nodes in the hidden layer immediately preceding the output layer of the neural network used in the previous learning is read. Next, proceed to step 103, acquire a new input value x, that is, a new value of a specific type of operating parameter related to the internal combustion engine, and store the new input value x, that is, a new value of a specific type of operating parameter related to the internal combustion engine in the electronic The storage part of the control unit 30 . And in step 103, the actual measurement value of the exhaust gas loss amount with respect to the new input value x is memorize|stored in the memory|storage part of the electronic control unit 30 as training data. That is, in step 103 , training data obtained by actual measurement of the newly acquired values of the operating parameters of the specific type related to the internal combustion engine are stored in the storage unit of the electronic control unit 30 .

Next, in step 104, it is determined whether or not the new input value x, that is, the newly acquired value of the operating parameter of the specific type related to the internal combustion engine, is between A and B indicating the preset range R, that is, the new input value. Is x greater than or equal to A and less than or equal to B. When the new input value x is between A and B representing the preset range R, the process proceeds to step 105, and the input value x, that is, the newly acquired value of the operating parameter of the specific type related to the internal combustion engine, is input to the neural network The node input of the layer is based on the output value y output from the node output of the neural network and the training data obtained by the actual measurement of the newly acquired values of the operating parameters of the specific category related to the internal combustion engine, using the error back propagation method, The weights of the neural network are learned in such a way that the difference between the output value y and the training data becomes small.

On the other hand, when it is determined in step 104 that the new input value x, that is, the newly acquired value of the operating parameter of the specific type related to the internal combustion engine is not between A and B indicating the preset range R, the process proceeds to step 106 , update the number K of nodes in the previous hidden layer of the output layer of the neural network, and increase the number K of nodes in the previous hidden layer of the output layer. At this time, in the first embodiment, the number K of nodes in the previous hidden layer of the output layer is increased by one. Next, in step 107 , the neural network is updated so that the number K of nodes in the hidden layer immediately preceding the output layer is increased, and the process proceeds to step 105 . In step 105, the training data newly obtained for the new input value x is also included in the training data, and the weight of the updated neural network is learned so that the difference between the output value y and the training data becomes smaller. That is, in step 105, the training data obtained by actual measurement of the newly acquired values of the operating parameters of the specific type related to the internal combustion engine and the training data of the operating parameters of the specific type related to the internal combustion engine within the preset range R are used. The training data obtained in advance by actual measurement, so that the output value y that changes according to the value of the specific type of operating parameter related to the internal combustion engine within the preset range and outside the preset range corresponds to the output value y. The weights of the updated neural network are learned so that the difference between the training data of the values of the specific types of operating parameters related to the internal combustion engine becomes smaller.

In this case, when the newly acquired value of the operating parameter of the specific type related to the internal combustion engine is outside the preset range, the value of the newly acquired operating parameter of the specific type related to the internal combustion engine may be determined by actual measurement. When the number of obtained training data is two or more than a certain number, the number of nodes in the hidden layer immediately preceding the output layer of the neural network is increased. Therefore, in the first embodiment, when the value of the newly acquired operating parameter of the specific type related to the internal combustion engine is out of the preset range, based on the newly acquired value of the specific type of operating parameter related to the internal combustion engine The number of nodes in the previous hidden layer of the output layer of the neural network is increased by increasing the number of training data obtained through actual measurement.

In addition, when there is a plurality of training data obtained by actual measurement for newly acquired values of operating parameters of a specific type related to the internal combustion engine that are outside the preset range, the difference between B and C in FIG. 8B may be determined. An increase in the data density of the training data within the range of the values of the preset operating parameters shown corresponds to an increase in the number of nodes in the hidden layer immediately preceding the output layer of the neural network. It should be noted that, in FIG. 8B , B and C respectively represent the minimum value and the maximum value of the range of the values of the preset operating parameters. Therefore, to be precise, it can be expressed by dividing the number of training data by The increase in data density obtained by the difference between the maximum value C and the minimum value B (C-B) of the preset range of the values of the operating parameters correspondingly increases the number of nodes in the hidden layer immediately preceding the output layer of the neural network. number increases.

As shown in FIG. 1 , the internal combustion engine used in the embodiment of the present invention includes an electronic control unit 30, and the electronic control unit 30 includes: a parameter value acquisition unit that acquires values of operating parameters of the internal combustion engine; The hidden layer and the neural network of the output layer perform operations; and a storage unit. Here, the input port 35 shown in FIG. 1 constitutes the above-mentioned parameter value acquisition unit, the CPU 34 constitutes the above-mentioned calculation unit, and the ROM 32 and the RAM 33 constitute the above-mentioned storage unit. Note that, in the CPU 34 , that is, the above-described calculation unit, the value of the operating parameter of the internal combustion engine is input to the input layer, and the output value that varies according to the value of the operating parameter of the internal combustion engine is output from the output layer. In addition, the range R set in advance with respect to the value of the operating parameter of the specific type related to the internal combustion engine is preliminarily stored in the ROM 32 , that is, the above-mentioned storage unit. In addition, the learned weights and the training data obtained in advance for the values of the specific types of operating parameters related to the internal combustion engine within the preset range R are stored in the RAM 33 , that is, the above-mentioned storage unit.

FIG. 11 shows a modification of the neural network used in the first embodiment of the present invention. In this modification, the output layer (L=4) has two nodes.

因此一方的节点处的δ^(L)的值成为δ^(L)＝y₁-y_t1，关于另一方的节点，利用y₂对平方和误差E进行偏微分

因此另一方的节点处的δ^(L)的值成为δ^(L)＝y₂-y_t12。关于输出层(L＝4)的各节点，当δ^(L)求出后，使用前式(6)求出前层的δ^(L-1)。这样，依次求出前层的δ，使用这些δ的值，根据前式(2)，关于各权重w求出误差函数E的微分即梯度

当求出梯度

后，使用该梯度

以使误差函数E的值减小的方式更新权重w的值。In this modification, as in the example shown in FIG. 9 , the input value x is set to the engine speed N (rpm). On the other hand, in this modification, one of the output amounts y ₁ is assumed to be the exhaust gas loss amount as in the example shown in FIG. 9 , and the other output amount y ₂ is assumed to be the quadratic of the input value x Some quantity of a function, such as a fuel consumption rate. Also in this modification, each node of the output layer (L=4) uses the identity function as the activation function. In the case where the output layer (L=4) has a plurality of nodes, as described above, using the square sum error E (the output values from each node are y ₁ , y ₂ . . . ) shown in the previous equation (8), The corresponding training data are y _t1 , y _t2 ...) as the error function E. In this case, as can be seen from the previous equation (7), with respect to one node, the sum of squares error E is partially differentiated by y ₁

Therefore, the value of δ (L ⁾ at one node becomes δ ^(L) = y ₁ -y _t1 , and with respect to the other node, the square sum error E is partially differentiated by y ₂

Therefore, the value of δ (L ⁾ at the other node becomes δ ^(L) =y ₂ -y _t12 . For each node of the output layer (L=4), after δ ^{(L) is} obtained, δ ^{(L-1) of the previous layer is obtained using the preceding equation (6)} . In this way, the δ of the previous layer is successively obtained, and using these values of δ, the gradient of the error function E, that is, the differential of the error function E, is obtained for each weight w according to the preceding formula (2).

When finding the gradient

After that, use the gradient

The value of the weight w is updated in such a way that the value of the error function E decreases.

Even when the output layer (L=4) of the neural network has a plurality of nodes as shown in FIG. 11, the output value z of the node passing through the hidden layer (L=3) preceding the output layer (L=4) ₁ and z ₂ are determined in the form of functions output from each node of the output layer (L=4). That is, what kind of function expression each output value y ₁ , y ₂ can represent is governed by the number of nodes in the hidden layer (L=3) immediately preceding the output layer (L=4). Therefore, in the neural network shown in Figure 11, when the number of nodes in the hidden layer is increased, as shown in Figure 11, the previous hidden layer (L=3) of the output layer (L=4) is increased. the number of nodes.

12 to 14B illustrate a second embodiment of the machine learning apparatus of the present invention. In the second embodiment, the operating parameters related to the internal combustion engine are composed of a plurality of categories of operating parameters, and the learning of the weights of the neural network is performed based on the values of the operating parameters related to the plurality of categories of the internal combustion engine. As a specific example, a case is shown in which the operating parameters of the internal combustion engine are composed of the engine speed, the accelerator opening (the amount of depression of the accelerator pedal), and the outside air temperature, and values based on the operating parameters of the internal combustion engine are created to estimate the output torque of the internal combustion engine. neural network model. In this specific example, as shown in FIG. 13 , the input layer (L=1) of the neural network is composed of three nodes, and an input value x ₁ representing the engine speed and an input value x representing the accelerator opening are input to each node. ₂ and the input value x ₃ representing the outside air temperature. In addition, the number of hidden layers (L=2, L=3) can be set to one or an arbitrary number, and the number of nodes of the hidden layers (L=2, L=3) can also be set to an arbitrary number number. It should be noted that, in the example shown in FIG. 13 , the number of nodes of the output layer (L=4) is set to one.

_On the other hand, in FIG. 14A, between A1 and B1, that is, R1, represents _a preset range _of the engine speed, and between _A2 and _B2 , that is, _R2 , represents a preset range of the accelerator opening. Between A3 and _B3 , that is, _R3 represents a preset range _of the outside air temperature. 14B is the same as FIG. 14A , between A ₁ and B ₁ represents a preset range of the engine speed, between A ₂ and B ₂ represents a preset range of the accelerator opening, and A ₃ Between B3 and _B3 indicates the preset range of the outside air temperature. It should be noted that, in the second embodiment, the accelerator opening is detected by the load sensor 41 , and the outside air temperature is detected by the atmospheric temperature sensor 24 . In addition, in the second embodiment, for example, the output torque of the internal combustion engine is actually measured by a torque sensor attached to the crankshaft of the internal combustion engine, and the torque obtained by the actual measurement is used as training data.

Also in this second embodiment, the output torque of the internal combustion engine actually measured for various input values x _n (n=1, 2, 3) within a preset range Rn is obtained as training data in advance, that is, Training data is obtained in advance for the values of the plurality of types of operating parameters related to the internal combustion engine within the preset range Rn by actual measurement, and is determined based on the values of the plurality of types of operating parameters related to the internal combustion engine and the training data. The structure of the neural network is to learn the weights of the neural network in advance so that the difference between the output value y and the training data corresponding to the values of the plurality of categories of operating parameters related to the internal combustion engine becomes small. The training data obtained in advance by actual measurement of the values of the plurality of types of operating parameters related to the internal combustion engine within the preset range Rn are stored in the storage unit of the electronic control unit 30 . Also in this second embodiment, a neural network having the same structure as the neural network used in the previous learning is used, and the weights of the neural network at the time of completion of the learning are used, and the learning is further performed on-board while the vehicle is running. FIG. 12 shows a learning processing routine of the second embodiment performed in the vehicle-mounted manner, and the learning processing routine is executed by interruption at regular intervals (for example, every one second).

Referring to FIG. 12 , first, in step 201 , the learned weights stored in the storage unit of the electronic control unit 30 and the training data used in the previous learning, that is, the internal combustion engine-related weights within the preset range Rn are read. The training data obtained in advance by actual measurement, the values An, Bn (n(n) representing the range of input data, that is, the preset range of the values of the operating parameters of the various categories related to the internal combustion engine. =1, 2, 3) (FIG. 13A). This learned weight is used as an initial value for the weight. Next, in step 202, the number K of nodes in the hidden layer immediately preceding the output layer of the neural network used in the previous learning is read. Next, the process proceeds to step 203, where a new input value x, that is, a new value of a plurality of categories of operating parameters related to the internal combustion engine is acquired, and the new input value x, that is, a new value of a plurality of categories of operating parameters related to the internal combustion engine, is stored in the storage unit of the electronic control unit 30 . Then, in step 203, the actual measurement value of the output torque of the internal combustion engine with respect to the new input value x is stored in the storage unit of the electronic control unit 30 as training data. That is, in step 203 , training data obtained by actual measurement of the newly acquired values of the operating parameters of a plurality of categories related to the internal combustion engine are stored in the storage unit of the electronic control unit 30 .

Next, in step 204, it is determined whether or not the new input value xn, that _is , the newly acquired values of the plurality of types of operating parameters related to the internal combustion engine, is within the preset range Rn (between An and Bn), that is, whether the new Is the input value x _n of , greater than or equal to An and less than or equal to Bn. When the new input value xn is within the preset range _Rn , the process proceeds to step 205, and each input value xn, that _is , the newly acquired values of the plurality of types of operating parameters related to the internal combustion engine, is sent to the input layer of the neural network. The corresponding node input is based on the output value y output from the node of the output layer of the neural network and the training data obtained by actual measurement of the newly acquired values of the operating parameters of multiple categories related to the internal combustion engine, using error back propagation method to learn the weights of the neural network in such a way that the difference between the output value y and the training data becomes smaller.

On the other hand, in step 204, it is determined that the new input value xn, that is, the value of at least one type of the operation parameter value of the plurality of types of operation parameter values related to the internal combustion engine newly acquired is not within the preset range _Rn (between An and Bn), for example, in FIG. 14B , the input value x ₁ representing the engine speed is within a predetermined range (B ₁ to C ₁ ) of B ₁ to C ₁ (B ₁ <C ₁ ) If the input value x ₃ representing the outside air temperature in FIG. 13B is within the preset range (C ₃ to A ₃ ) of C ₃ to A ₃ (C ₃ <A ₃ ), the process proceeds to step 206 . In _{step 206} , _first , the _{density D} ( ₌ Number of training data/(C _n -B _n ) or number of training data/(A _n -C _n )).

In FIG. 14B , B ₁ and C ₁ respectively represent the minimum and maximum values of the preset range of the engine speed, that is, the minimum and maximum values of the preset ranges of the values of the operating parameters, and the training data density D represents the A value obtained by dividing the number of training data by the difference (C ₁ -B ₁ ) between the maximum value C ₁ and the minimum value B ₁ in a predetermined range representing the values of the operating parameters. In addition, in FIG. 14B , C ₃ and A ₃ respectively indicate the minimum and maximum values of the range of the preset outside air temperature, that is, the minimum and maximum values of the preset ranges of the values of the operating parameters, and the training data density. D represents a value obtained by dividing the number of training data by the difference (C ₃ -A ₃ ) between the maximum value C ₃ and the minimum value A ₃ in a predetermined range representing the value of the operating parameter. In step 206, after the training data density D is calculated, it is determined whether the training data density D becomes higher than the predetermined data density D ₀ . In the event that the training data density D is lower than the predetermined data density D ₀ , the processing loop is completed.

On the other hand, when it is determined in step 206 that the training data density D has become higher than the predetermined data density D ₀ , the process proceeds to step 207 . In this case, when D(=number of training data/(A _n −C _n ))>D ₀ , the number of additional nodes α is calculated by the following equation.

Number of additional nodes α=round{(K/(Bn-An))·(An-Cn)}

On the other hand, when D (=number of training data/(C _n −B _n ))>D ₀ , the number of additional nodes α is calculated by the following equation.

Number of additional nodes α=round{(K/(Bn-An))·(Cn-Bn)}

It should be noted that, in the above formula, K represents the number of nodes, and round means rounding.

After calculating the number of additional nodes α in step 207, go to step 208 to update the number K of nodes in the previous hidden layer of the output layer of the neural network, so that the number K of nodes in the previous hidden layer of the output layer is increased. The number of nodes α (K←K+α) is added. In this way, in the second embodiment, when the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value of the preset range representing the value of the operating parameter increases, the data density increases. The number of nodes in the previous hidden layer of the output layer of the neural network. That is, in the second embodiment, the data density is increased in accordance with the increase in the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value of the preset range representing the value of the operating parameter. Increase the number of nodes in the previous hidden layer of the output layer of the neural network.

On the other hand, as described above, when the training data density D reaches the predetermined data density D ₀ from the step 206 to the step 207, the (A _n − The value of C _n ) and the value of (C _n -B _n ) are proportional to the number of training data. Therefore, as can be seen from the above formula, the number of additional nodes α is proportional to the range (B _n to C _n ) to which the new input value x _n belongs or the number of training data in the range (C _n to A _n ). That is, in the second embodiment, the output layer of the neural network is increased in accordance with the increase in the number of training data obtained by actual measurement of newly acquired values of operating parameters of a plurality of categories related to the internal combustion engine. The number of nodes in the previous hidden layer.

When the number K of nodes in the previous hidden layer of the output layer is increased by the number of additional nodes α in step 208 (K←K+α), go to step 209, so that the number of nodes in the previous hidden layer of the output layer is increased. The neural network is updated by increasing the number K. Next, go to step 205 . In step 205, the training data newly obtained for the new input value x is also included in the training data, and the weight of the updated neural network is learned so that the difference between the output value y and the training data becomes smaller. That is, in step 205, the training data obtained by actual measurement of the newly acquired values of the operating parameters of the plurality of categories related to the internal combustion engine and the operation of the plurality of categories related to the internal combustion engine within the preset range Rn are used. The training data for which parameter values are obtained in advance through actual measurement, and output values y that vary according to the values of various types of operating parameters related to the internal combustion engine within a preset range and outside the preset range correspond to The weights of the updated neural network are learned so that the difference between the training data of the values of the operating parameters of the plurality of categories related to the internal combustion engine becomes smaller.

In the second embodiment of the present invention, the values An and Bn representing the range Rn preset for the values of the plurality of types of operating parameters related to the internal combustion engine are prestored in the ROM 32 , that is, the above-mentioned storage unit. In addition, the learned weights and the training data obtained in advance for the values of the operating parameters of the plurality of categories related to the internal combustion engine within the preset range Rn are stored in the RAM 33 , that is, the above-mentioned storage unit.

FIG. 15 shows a modification of the neural network used in the second embodiment of the present invention. In this modification, the output layer (L=4) has two nodes.

In this modification, as in the example shown in FIG. 13 , the input value x ₁ is the engine speed, the input value x ₂ is the accelerator opening degree, and the input value x ₃ is the outside air temperature. On the other hand, in this modification, one of the output amounts y ₁ is assumed to be the output torque of the internal combustion engine, and the other output amount y ₂ is assumed to be the thermal efficiency of the internal combustion engine, as in the example shown in FIG. 13 . This thermal efficiency is calculated based on the detected values of the engine speed, engine load, intake air pressure, intake air temperature, exhaust gas pressure, exhaust gas temperature, engine cooling water temperature, etc. when the internal combustion engine is actually operated. Therefore, the thermal efficiency represents a value obtained by actual measurement. . Also in this modification, each node of the output layer (L=4) uses the identity function as the activation function. Furthermore, in this modification example, the error back propagation method is used by using the square sum error E shown in the preceding equation (8) as the training data with the actual measured value of the output torque of the internal combustion engine and the actual measured value of the thermal efficiency as the training data. , update the value of the weight w in such a way that the value of the error function E decreases.

Even when the output layer (L=4) of the neural network has a plurality of nodes as shown in FIG. 15, the output value z of the node passing through the hidden layer (L=3) preceding the output layer (L=4) ₁ , z ₂ , z ₃ , and z ₄ are determined in the form of functions output from each node of the output layer (L=4). That is, what kind of function expression each output value y ₁ , y ₂ can represent is governed by the number of nodes in the hidden layer (L=3) immediately preceding the output layer (L=4). Therefore, in the neural network shown in FIG. 15 , when the number of nodes in the hidden layer is increased, the number of nodes in the hidden layer (L=3) immediately preceding the output layer (L=4) is increased.

16 and 17 show a third embodiment of the machine learning apparatus of the present invention. Also in this third embodiment, the operating parameters related to the internal combustion engine include a plurality of categories of operating parameters, and the learning of the weight of the neural network is performed based on the values of the operating parameters of the plural categories related to the internal combustion engine. Also in this third embodiment, a range of values of the operating parameters of each category is set in advance for each of a plurality of categories of operating parameters related to the internal combustion engine. FIG. 17 shows, as an example, a case where the operating parameters related to the internal combustion engine are composed of two types of operating parameters. In FIG. 17 , the preset range of the values of the operating parameters of one type is represented by Rx, and the other type is represented by Rx. The preset range of the value of the operating parameter is represented by Ry. In the third embodiment, as shown in FIG. 17 , the preset ranges Rx and Ry for the values of the operating parameters of each category are divided into a plurality of groups, and the ranges Rx and Ry that pass the values of the operating parameters of each category are preset. A plurality of divided regions [Xn, Ym] (n=1, 2...n, m=1, 2...m) defined by the combination of the divided division ranges.

In addition, in FIG. 17, X1, X2...Xn and Y1, Y2...Yn each show the division|segmentation range of the value of each type of operation parameter. In the third embodiment, as a specific example, the operating parameters of the internal combustion engine are composed of the engine speed and the outside air temperature, and values based on the operating parameters of the internal combustion engine are created to estimate the HC discharge amount from the internal combustion engine. neural network model. In this case, X1, X2,...Xn represent, for example, the rotational speed of the internal combustion engine divided at every 1000rpm (1000rpm≤X1<200rpm, 2000rpm≤X2<3000rpm...), and Y1, Y2...Yn are, for example, the external engine speed divided at every 10°C. Air temperature (-30℃≤Y1<-20℃, -20℃≤Y2<-10℃…).

In the third embodiment, an independent neural network is created for each of the divided regions [Xn, Ym]. In these neural networks, the input layer (L=1) is composed of two nodes, and an input value x1 representing the engine speed and an input value x2 representing the outside air temperature are input to each node of the input layer (L=1). In addition, the number of hidden layers (L=2, L=3) can be set to one or an arbitrary number, and the number of nodes of the hidden layers (L=2, L=3) can also be set to an arbitrary number number. It should be noted that, in any neural network, the number of nodes of the output layer (L=4) is set to one. It should be noted that, also in the third embodiment, as a modification, the number of nodes of the output layer (L=4) may be set to two. In this case, for example, the output amount from one node of the output layer (L=4) is set to the HC discharge amount from the internal combustion engine, and the output amount from the other node of the output layer (L=4) is set to is the _NOx emission from the internal combustion engine.

In this third embodiment, the number of nodes in the hidden layer (L=3) is different for each neural network. Hereinafter, the output layer of the neural network in the divided region [Xn, Ym] will be divided into the previous hidden layer of the neural network. The number of nodes is represented by Knm. The number Knm of nodes of the hidden layer is set in advance according to the complexity of the change of the training data in each of the divided regions [Xn, Ym] with respect to the change of the input value. In addition, in this 3rd Example, the HC sensor was arrange|positioned in the exhaust passage instead of the _NOx sensor 23 shown in FIG. In the third embodiment, the HC discharge amount from the internal combustion engine is actually measured by the HC sensor, and the HC discharge amount obtained by the actual measurement is used as training data. In addition, in the modification of the above-mentioned third embodiment, the HC sensor is arranged in the exhaust passage in addition to the _NOx sensor 23 shown in FIG. 1 .

In the third embodiment, each of the divided regions [Xn, Ym] (n=1, 2 . , m=1, 2… The values of the operating parameters of each category are obtained in advance through actual measurement to obtain training data, and based on the values of the operating parameters of the plurality of categories related to the internal combustion engine and the training data, the number Knm of nodes including the hidden layer is also determined. The structure of the neural network that divides the regions [Xn, Ym] is learned in advance for each of the divided regions [Xn , Ym] the weights of the neural network. Therefore, in the third embodiment, the previously learned division region [Xn, Ym] (n=1, 2...n, m=1, 2...m) is sometimes referred to as a learned division hereinafter. Area [Xn, Ym]. It should be noted that the training data obtained in advance by actual measurement of the values of the plurality of types of operating parameters related to the internal combustion engine within the preset ranges Rx and Ry are stored in the storage unit of the electronic control unit 30 . Also in this third embodiment, a neural network having the same structure as the neural network used in the previous learning is used for each of the divided regions [Xn, Ym], and the weight of the neural network at the time of completion of the learning is used, and the weight of the neural network at the time of completion of the learning is used. In-vehicle mode for further learning. FIG. 16 shows a learning processing routine of the third embodiment performed in a vehicle-mounted manner, and the learning processing routine is executed by interruption at regular intervals (for example, every one second).

Referring to FIG. 16 , first, in step 301, the learned weights stored in the storage unit of the electronic control unit 30 and the training data used in the previous learning, that is, the sum of the preset ranges Rx and Ry, are read. The training data obtained in advance by actual measurement for the values of the operating parameters of a plurality of categories related to the internal combustion engine, and each learned divided area [Xn, Ym] (n=1, 2...n, m=1, 2...m). This learned weight is used as an initial value for the weight. Next, in step 302, the number Knm of nodes in the hidden layer immediately preceding the output layer used in the previous learning for each learned partition region [Xn, Ym] is read. Next, the process proceeds to step 303, and new input values x1, x2, that is, new values of the operation parameters of the plurality of categories related to the internal combustion engine are acquired, and the new input values x1, x2 are the new operation parameters of the plurality of categories related to the internal combustion engine. The value of the parameter is stored in the storage unit of the electronic control unit 30 . And in step 303, the actual measurement value of HC discharge|emission amount with respect to the new input value x1, x2 is memorize|stored in the memory|storage part of the electronic control unit 30 as training data. That is, in step 303 , training data obtained by actual measurement of the newly acquired values of a plurality of types of operating parameters related to the internal combustion engine are stored in the storage unit of the electronic control unit 30 .

Next, in step 304, it is determined whether or not the new input values x1 and x2 are within the learned division area [Xn, Ym], that is, whether or not the newly acquired values of the operating parameters of a plurality of categories related to the internal combustion engine are within the preset values within the range of Rx and Ry. When the new input values x1 and x2 are within the learned division region [Xn, Ym], that is, when the newly acquired values of the operating parameters of a plurality of categories related to the internal combustion engine are within the preset ranges Rx and Ry , proceed to step 305, and transfer the input values x1, x2, that is, the newly acquired values of the operating parameters of multiple categories related to the internal combustion engine, to the learned division area to which the newly acquired values of the operating parameters of multiple categories related to the internal combustion engine belong [ The input to each node of the input layer of the neural network of Xn, Ym] is obtained by actual measurement based on the output value y output from the node of the output layer of the neural network and the newly acquired values of the operating parameters of the plurality of categories related to the internal combustion engine. The obtained training data, using the error back-propagation method, further learns the learned division area to which the values of the newly acquired operating parameters of multiple categories related to the internal combustion engine belong in a way to make the difference between the output value y and the training data smaller [ Xn, Ym] the weights of the neural network.

On the other hand, when it is determined in step 304 that the new input values x1 and x2 are not within the learned division area [Xn, Ym], for example, in FIG. 17 , the input values x1 and x2 belong to the input values x1 and x2 The unlearned area [Xa, Yb] delimited by the combination of the preset ranges is set outside the learned divided area [Xn, Ym]. That is, in other words, when the value of at least one of the newly acquired operating parameters of the internal combustion engine is outside the preset ranges Rx and Ry, the value of the operating parameter of each category belongs to and passes through The unlearned region [Xa, Yb] defined by the combination of the preset ranges of the values of the operating parameters of each category is set outside the preset ranges Rx, Ry.

The example shown in FIG. 17 shows the case where the newly acquired value of the operating parameter of one category related to the internal combustion engine is outside the preset range Rx, and the value of the operating parameter of another category related to the internal combustion engine is newly acquired It belongs to the division range Y2 within the preset range Ry, and in this case, the unlearned area [Xa, Yb] is outside the preset range Rx for the value of the operating parameter of one category and is in the other category. Within the division range Y2 to which the value of the operating parameter belongs, it is set adjacent to the learned division area [Xn, Ym] in the division range Y2.

When the unlearned area [Xa, Yb] is set, the process proceeds to step 306 . In step 306, first, the training data density D in the unlearned area [Xa, Yb] to which the new input values x1 and x2 belong is calculated. The training data density D (=the number of training data/[Xa, Yb]) represents a preset value of the operating parameter value of each category divided by the number of training data divided by the area of the unlearned region [Xa, Yb] The value obtained by multiplying the ranges. Next, it is determined whether the training data density D becomes higher than the predetermined data density D ₀ and whether the variance S ² of the training data in the new area [Xa, Yb] to which the new input values x1 and x2 belong is higher than the predetermined variance S ²⁰ _large . In the case where the training data density D is lower than the predetermined data density D ₀ or in the case where the variance S ² of the training data is smaller than the predetermined variance S ² ₀ , the processing loop is completed.

On the other hand, when it is determined in step 306 that the training data density D is higher than the predetermined data density D ₀ and the variance S ² of the training data is larger than the predetermined variance S ² ₀ , the process proceeds to step 307 . It should be noted that, in step 306, the determination of whether the variance S ² is greater than the predetermined variance S ² ₀ may be omitted, and only the training data density D becomes higher than the predetermined data density D ₀ . In this case, when the training data density D is lower than the predetermined data density D ₀ , the processing loop is completed, and when it is determined that the training data density D is higher than the predetermined data density D ₀ , the process proceeds to step 307 . In step 307, based on the following formula for calculating the number of nodes, the relative value to the unlearned area [ Xa, Yb] the number of nodes Kab.

Number of nodes Kab=1/NΣΣKij (i=(a-1)～(a+1), j=(b-1)～(b+1))

It should be noted that, in the above formula, N represents the number of learned divided regions [Xn, Ym] that exist adjacent to the unlearned region [Xa, Yb]. In this case, there is an unused divided area [Xn, Ym] in the adjacent divided areas [Xn, Ym] around the unlearned area [Xa, Yb], that is, there is no divided area [Xn with the number of nodes Knm] , Ym], the division area [Xn, Ym] is excluded from the calculation of the number N. For example, taking the example shown in FIG. 15, the number of nodes Kn1 of the adjacent learned divided areas [Xn, Y1] around the unlearned area [Xa, Yb], and the learned divided areas [Xn, Y2] The average value of the number of nodes Kn2 of , and the number of nodes Kn3 of the learned divided area [Xn, Y3] is set to Kab with respect to the number of nodes of the unlearned area [Xa, Yb].

When the relationship between the change of the training data and the change of the input value in each of the divided regions [Xn, Ym] is simple, sufficient learning can be performed even if the number Knm of nodes in the hidden layer is small, but in each of the divided regions When the relationship between changes in training data in [Xn, Ym] with respect to changes in input values is complicated, sufficient learning cannot be performed unless the number Knm of nodes in the hidden layer is increased. Therefore, as described above, the number of nodes Knm of the hidden layer nodes of the neural network in the learned division area [Xn, Ym] is relative to the input value according to the change of the training data in each learned division area [Xn, Ym] The complexity of the change is set. In the case where the two divided areas [Xn, Ym] are close, the relationship between the change of the training data and the change of the input value is similar between these divided areas [Xn, Ym]. Therefore, between the two divided areas [Xn, Ym] , Ym] are close to each other, the same number can be used as the number Knm of nodes of the hidden layer. Therefore, in this third embodiment, the average value of the number of nodes Knm in the adjacent learned divided regions [Xn, Ym] around the unlearned region [Xa, Yb] is set relative to the unlearned region [Xa , Yb] the number of nodes Kab.

Here, as a modification of the third embodiment, a method of obtaining the number of nodes Kab relative to the unlearned area [Xa, Yb] considering the number of training data in the unlearned area [Xa, Yb] is considered Give a brief explanation. That is, when the number of training data in the unlearned region [Xa, Yb] is larger than the number of training data in the adjacent learned divided regions [Xn, Ym] around the unlearned region [Xa, Yb] , the number of nodes Kab relative to the unlearned area [Xa, Yb] is preferably larger than the number of nodes Knm in the adjacent learned divided areas [Xn, Ym] around the unlearned area [Xa, Yb]. Therefore, in this modification, the average value MD of the number of training data in the adjacent learned regions [Xn, Ym] around the unlearned region [Xa, Yb] is obtained, and the unlearned region [Xa, Ym] is calculated by dividing the unlearned region [Xa, Ym] The number of input data MN in Yb] is divided by the average value MD to obtain the node number increase rate RK (=MN/MD). ] is multiplied by the node number increase rate RK to obtain the final node number Kab with respect to the unlearned area [Xa, Yb].

After calculating the number of nodes Kab with respect to the unlearned area [Xa, Yb] in step 307, the process proceeds to step 308, and a new neural network with respect to the unlearned area [Xa, Yb] is created. In this new neural network, the number of nodes is set to two for the input layer, Kab for the previous hidden layer of the output layer, and one or more for the output layer. Next, go to step 305 . In step 305, regarding the unlearned area [Xa, Yb], the weight of the neural network created for the unlearned area [Xa, Yb] is learned so that the difference between the output value y and the training data becomes small.

Next, a specific example of a case where the machine learning apparatus of the present invention is applied to a special low-load internal combustion engine will be described with reference to FIGS. 18A to 19B . In this specific example, as shown in FIG. 12 , a neural network having 4 nodes in the hidden layer (L=3) is used to create an output representing the _NOx emission amount according to the opening degree of the throttle valve 12 , the engine speed and the ignition timing. The output value y of the model. It should be noted that, in the internal combustion engine used in this specific example, the use range of the opening degree of the throttle valve 12 is set to 5.5° to 11.5° (the opening degree of the throttle valve 12 at the maximum valve closing position is set to 0 °), the use range of the engine speed is set between 1600 (rpm) and 3000 (rpm), and the use range of the ignition timing is set from 0° (compression top dead center) to ATDC (compression top dead center) point) between 40°.

FIG. 18A shows the distribution of the training data with respect to the ignition timing and the engine speed, and FIG. 18B shows the distribution of the training data with respect to the throttle valve opening and the ignition timing. In addition, in FIG. 18A and FIG. 18B, the black circle shows the place where the training data acquired in advance exists, and the triangle mark shows the place where the training data is not acquired beforehand. It can be seen from FIGS. 18A and 18B that the training data has been acquired in advance for which throttle valve opening, which engine speed, and which ignition timing. For example, in FIG. 18A , when the engine speed N is 2000 (rpm) and the ignition timing is ATDC 20°, the training data is acquired in advance. As shown in FIG. 18B , when the ignition timing is ATDC 20°, for each There are various kinds of throttle opening, and the training data is obtained in advance.

On the other hand, in this specific example, the throttle valve opening, the engine speed, and the ignition timing are input to each node of the input layer (L=1) of the neural network, so that the output value y and the output value y indicated by the _NOx sensor 23 The weights of the neural network are learned such that the difference between the detected _NOx emission and the training data becomes smaller. The relationship between the learned output value y and the training data is shown in FIGS. 18C , 19A and 19B . It should be noted that in FIGS. 18C , 19A and 19B , the learned output value y and the value of the training data are equal to each other. It is shown normalized so that the maximum value becomes 1.

As described above, in the internal combustion engine used in this specific example, the use range of the opening degree of the throttle valve 12 is set to be between 5.5° and 11.5°, and the use range of the engine speed N is set to be between 1600 (rpm) and 1600 (rpm). Between 3000 (rpm), the use range of the ignition timing is set between 0° (compression top dead center) and ATDC40°. In FIG. 18C , the _NOx emission amount using the throttle valve opening, the engine speed N, and the ignition timing within these use ranges is acquired as training data in advance, so that the difference between the output value y and the previously acquired training data is obtained. The relationship between the learned output value y and the training data when the weight of the neural network has been learned in a smaller manner is shown by a circle.

As shown in FIG. 18C , the circle marks representing the relationship between the output value y after learning and the training data are concentrated on a straight line, so it can be seen that the output value y after learning agrees with the training data. For example, taking the opening degree of the throttle valve 12 as an example, the opening degree of the throttle valve 12 may deviate from the standard opening degree due to individual differences of the engine or changes over time, even if the usage range of the opening degree of the throttle valve 12 is set It is set to be between 5.5° and 11.5°. Actually, the opening degree of the throttle valve 12 may exceed the preset usage range. The triangular mark shown in FIGS. 18A and 18B indicates a place where training data is newly acquired when the opening degree of the throttle valve 12 exceeds a preset usage range and becomes 13.5°.

The triangular mark in FIG. 18C shows that when the opening degree of the throttle valve 12 exceeds the preset usage range and becomes 13.5° in this way, the weight of the neural network is performed not using the newly acquired training data but only using the previously acquired training data situation of learning. In this case, it can be seen that the estimated value of the NO _X emission amount when the opening degree of the throttle valve 12 exceeds the preset use range and becomes 13.5° greatly deviates from the actual measurement value. On the other hand, the circle mark in FIG. 19A shows the use of both the new training data acquired when the opening degree of the throttle valve 12 exceeds the preset usage range and becomes 13.5° and the previously acquired training data. The training data is used to learn the weights of the neural network. In this case, it can be seen that the estimated value of the _NOx emission amount deviates from the actual measured value as a whole.

On the other hand, the circle mark in FIG. 19B shows the use of new training data and previously acquired training obtained when the opening degree of the throttle valve 12 exceeds the preset usage range and becomes 13.5°, as in FIG. 19A . The training data of both the data is different from the case in which the weight of the neural network is learned after the number of nodes in the hidden layer (L=3) of the neural network is increased from four to seven, different from FIG. 19A . In this case, it can be seen that the estimated value of the NO _X emission amount and the actual measurement value agree with high accuracy. In this way, when the value of the newly acquired operating parameter of the internal combustion engine is outside the preset range, the estimation accuracy can be improved by increasing the number of nodes in the hidden layer immediately preceding the output layer of the neural network.

20 to 25 show a fourth example of a case where the machine learning apparatus of the present invention is applied to automatic adjustment of an air conditioner (air conditioner). In this embodiment, the optimal air volume, air direction, and operation time of the air conditioner are automatically set according to the temperature, humidity, location, and the size of the room in which the air conditioner is installed. In this case, the conditions for the use of the air conditioner and the range of locations, that is, the range of use of the values of the operating parameters such as temperature, humidity, location, and the size of the room where the air conditioner is installed, can be preconceived according to the type of the air conditioner. The predetermined range of the values of the operating parameters of the air conditioner is to learn the weights of the neural network in advance so that the difference between the output value of the neural network and the optimal air volume, air direction, and operation time of the air conditioner becomes small.

However, in this case, the value of the operating parameter of the air conditioner may be outside the preset range. In this case, since learning based on actual values is not performed outside the preset range, neural Since the output value calculated by the network will be a value that deviates greatly from the actual value, in this embodiment also, when the value of the newly acquired operating parameter related to the air conditioner is outside the preset range, the neural network is set to The number of nodes in the previous hidden layer of the output layer is increased, or the number of neural networks is increased, and the training data obtained from the newly acquired values of the operating parameters related to the air conditioner and the preset range are used. The weights of the neural network are learned from the training data obtained by the values of the operating parameters related to the air conditioner.

Next, the fourth embodiment will be specifically described. 20, 50 denotes an air conditioner main body, 51 denotes a blower motor disposed in the air conditioner main body 50, 52 denotes an air direction adjusting motor disposed in the air conditioner main body 50, 53 denotes a thermometer for detecting the air temperature, and 54 denotes a thermometer for detecting the atmosphere A hygrometer of humidity, 55 denotes a GPS for detecting the installation position of the air conditioner, and 56 denotes an electronic control unit having the same structure as that of the electronic control unit 30 shown in FIG. 1 . As shown in FIG. 20, the temperature detected by the thermometer 53, the humidity of the atmosphere detected by the hygrometer 54, and the position information detected by the GPS 55 are input to the electronic control unit 56, and output from the electronic control unit 56 is used to obtain the optimum The drive signal of the blower motor 51 for the air volume of the air conditioner and the drive signal of the wind direction adjustment motor 52 for obtaining the optimum air direction of the air conditioner. In addition, the size of the room in which the air conditioner is installed is manually input to the electronic control unit 56, for example.

FIG. 21 shows the neural network used in this fourth embodiment. In this fourth embodiment, as shown in FIG. 21 , the input layer (L=1) of the neural network consists of four nodes, and input values x ₁ representing air temperature, input values x ₂ representing humidity, The input value x ₃ representing the position and the input value x ₄ representing the size of the room where the air conditioner is installed are represented. In addition, the number of hidden layers (L=2, L=3) can be set to one or an arbitrary number, and the number of nodes of the hidden layers (L=2, L=3) can also be set to any number number. In addition, in the fourth embodiment, the output layer (L=4) is composed of three nodes, and the output value y ₁ representing the air volume of the air conditioner, the output value y 2 representing the air direction of the air conditioner, and the output value y ₂ representing the air conditioner are output from each node. The input value y ₃ of the operating time.

_On the other hand, in FIG. 22A, between A1 and B1, that is, R1, represents _a preset range _of air temperature (for example, -5°C to 40°C ₎ , and between A2 and _B2 , that is, _R2 , represents humidity. The pre-set range (for example, 30% to 90%), between A ₃ and B ₃ , that is, R ₃ represents the pre-set range of the position (for example, between 20 degrees to 46 degrees north latitude), A ₄ and B Between B ₄ ie R ₄ represents a preset range of the size of the room where the air conditioner is installed. 22B is the same as FIG. 22A , between A ₁ to B ₁ shows a preset range of air temperature, between A ₂ and B ₂ shows a preset range of humidity, A ₃ and B ₃ Between represents the preset range of the position, and between A ₄ and B ₄ represents the preset range of the size of the room where the air conditioner is installed.

Also in this fourth embodiment, the air volume, air direction and operation time of the optimal air conditioner actually measured for various input values x _n (n=1, 2, 3, 4) within the preset range Rn are used as The training data is obtained in advance by training data, that is, the values of the operating parameters of the plurality of categories related to the air conditioner within the preset range Rn are obtained in advance through actual measurement. The values of the operating parameters and the training data determine the structure of the neural network so that the difference between the output values y ₁ , y ₂ , and y ₃ and the training data corresponding to the values of the operating parameters of a plurality of categories related to the air conditioner becomes small. way to learn the weights of the neural network in advance. The training data obtained in advance by actual measurement of the values of the plurality of types of operating parameters related to the air conditioner within the preset range Rn are stored in the storage unit of the electronic control unit 56 .

Also in this fourth embodiment, a neural network having the same structure as the neural network used in the previous learning is used, and the weight of the neural network at the time of completion of the learning is used, and the learning is further performed on-board while the vehicle is running. FIG. 23 shows a learning processing routine of the fourth embodiment carried out in the vehicle-mounted manner, and the learning processing routine is executed by interruption at regular intervals (for example, every one second). It should be noted that the processing performed in each step of the learning processing routine shown in FIG. 23 is different from the type of input value and the number of input values, and the type of output value and the number of output values. The processing is the same as that performed in each step of the learning processing routine shown in FIG. 12 .

23 , first, in step 401, the learned weights stored in the storage unit of the electronic control unit 56 and the training data used in the previous learning, that is, the sum of the weights within the preset range Rn, are read. Training data obtained in advance by actual measurement for the values of the operating parameters of the plurality of categories related to the air conditioner, and values An and Bn indicating the range of the input data, that is, the preset range of the values of the operating parameters of the plurality of categories related to the air conditioner (n=1, 2, 3, 4) (FIG. 22A). This learned weight is used as an initial value for the weight. Next, in step 402, the number K of nodes in the hidden layer immediately preceding the output layer of the neural network used in the previous learning is read. Next, the process proceeds to step 403, where a new input value x, that is, new values of operating parameters of a plurality of categories related to the air conditioner is acquired, and the new input value x, that is, a new value of the operating parameters of a plurality of categories related to the air conditioner, is stored in the storage unit of the electronic control unit 56 . Then, in step 403, the actual measurement values of the air volume, air direction, and operation time of the air conditioner with respect to the new input value x are stored in the storage unit of the electronic control unit 56 as training data. That is, in step 403 , training data obtained by actual measurement of the newly acquired values of the operating parameters of the plurality of categories related to the air conditioner are stored in the storage unit of the electronic control unit 56 .

Next, in step 404, it is determined whether the new input value xn, that _is , the newly acquired values of the operating parameters of a plurality of categories related to the air conditioner, is within the preset range Rn (between An and Bn), that is, whether the new Is the input value x _n greater than or equal to An and less than or equal to Bn. When the new input value xn is within the preset range _Rn , the process proceeds to step 405, and each input value xn, that _is , the newly acquired values of the operating parameters of the plurality of categories related to the air conditioner, is sent to the input layer of the neural network. The corresponding node input is based on the output values y ₁ , y ₂ , and y ₃ output from the nodes of the output layer of the neural network, and the training obtained by the actual measurement of the newly acquired values of the operating parameters of a plurality of categories related to the air conditioner The weights of the neural network are learned in such a way that the difference between the output values y ₁ , y ₂ , and y ₃ and the training data becomes smaller by using the error back-propagation method.

On the other hand, it is determined in step 404 that the new input value xn, that is, the value of the operating parameter of at least one category among the values of the operating parameters of the plurality of categories newly acquired related to the air conditioner is not within the preset range _Rn (between An and Bn), for example, in FIG. 22B , the input value x ₁ representing the air temperature is within a predetermined range (B ₁ to C ₁ ) of B ₁ to C ₁ (B ₁ <C ₁ ) 22B , or when the input value x ₃ representing the position is within a preset range (C ₃ to A ₃ ) of C ₃ to A ₃ (C ₃ <A ₃ ), the process proceeds to step 406 . In _{step 406} , _first , the _{density D} ( ₌ Number of training data/(C _n -B _n ) or number of training data/(A _n -C _n )). The definition of the training data density D is as described above. In step 406, after the training data density D is calculated, it is determined whether the training data density D becomes higher than the predetermined data density D ₀ . In the event that the training data density D is lower than the predetermined data density D ₀ , the processing loop is completed.

On the other hand, when it is determined in step 406 that the training data density D has become higher than the predetermined data density D ₀ , the process proceeds to step 407 . In this case, when D(=number of training data/(A _n -C _n ))>D ₀ , the number of additional nodes α is calculated by the following equation.

Number of additional nodes α=round{(K/(Bn-An))·(An-Cn)}

On the other hand, when D (=number of training data/(C _n -B _n ))>D ₀ , the number of additional nodes α is calculated by the following equation.

Number of additional nodes α=round{(K/(Bn-An))·(Cn-Bn)}

It should be noted that, in the above formula, K represents the number of nodes, and round means rounding.

After calculating the number of additional nodes α in step 407, go to step 408, update the number K of nodes in the previous hidden layer of the output layer of the neural network, and increase the number K of nodes in the previous hidden layer of the output layer The number of nodes α (K←K+α) is added. In this way, in the fourth embodiment, when the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value of the preset range representing the value of the operating parameter increases, the data density increases. The number of nodes in the previous hidden layer of the output layer of the neural network. That is, in the fourth embodiment, the data density is increased in accordance with the increase in the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value of the preset range of values representing the operating parameters. Increase the number of nodes in the previous hidden layer of the output layer of the neural network.

When the number K of nodes in the previous hidden layer of the output layer is increased by the number of additional nodes α in step 408 (K←K+α), go to step 409, so that the number of nodes in the previous hidden layer of the output layer is increased. The neural network is updated by increasing the number K. Next, go to step 405 . In step 405, the training data newly obtained for the new input value x is also included in the training data, so that the difference between the output values y ₁ , y ₂ , y ₃ and the training data becomes smaller to learn the updated neural network the weight of. That is, in step 405, the training data obtained by actual measurement of the newly acquired values of the operating parameters of the plurality of categories related to the air conditioner and the operation of the plurality of categories related to the air conditioner within the preset range Rn are used. The training data for which the parameter values are obtained in advance by actual measurement, and the output values y ₁ , y 1 , and The weights of the updated neural network are learned so that the difference between y ₂ and y ₃ and the training data corresponding to the values of the operating parameters of the plurality of categories related to the air conditioner becomes smaller.

24 and 25 show a modification of the fourth embodiment. In this modification, the preset range of the value of each type of operating parameter related to the air conditioner is divided into a plurality of groups. That is, Rw, Rx, Ry, and Rz in FIG. 24 represent the preset ranges of temperature, humidity, position, and the size of the room where the air conditioner is installed, respectively. As shown in FIG. The predetermined range of the size of the air-conditioned room is divided into a plurality of parts. It should _be noted that, in FIG. ₂₄ , W ₁ , _{W 2} . . . _{W n} , _{X 1} , _{X 2 .} The division range of the value of the operating parameter.

Furthermore, in this modification, a plurality of divided regions [Wi, Xj, Yk, Zl] (i=1) demarcated by the combination of the divided ranges obtained by dividing the values of the operating parameters of each category are preset. , 2...n, j=1, 2...n, k=1, 2...n, l=1, 2...n), an independent neural network is created for each divided area [Wi, Xj, Yk, Zl] . These neural networks have the configuration shown in FIG. 22 . In this case, the number of nodes in the hidden layer (L=3) is different for each neural network. Hereinafter, the hidden layer before the output layer of the neural network in the region [Wi, Xj, Yk, Zl] will be divided The number of nodes is represented by Ki, j, k, l. The numbers Ki, j, k, and l of nodes in the hidden layer are set in advance according to the complexity of the change of the training data in each of the divided regions [Wi, Xj, Yk, Zl] with respect to the change of the input value.

In this modification, various input values x1, x2, x3, and x4, ie, air temperatures, are applied to each of the divided regions [Wi, Xj, Yk, Zl] formed in the preset ranges Rw, Rx, Ry, and Rz. , humidity, location, and the size of the room where the air conditioner is installed. The measured air volume, wind direction, and operating time of the air conditioner are obtained as training data in advance. The values of the operating parameters of the plurality of categories related to the air conditioner are obtained in advance by actual measurement, and the training data is obtained in advance. Based on the values of the operating parameters of the plurality of categories related to the air conditioner and the training data, the numbers Ki and j of the nodes in the hidden layer are also included. , k, and l to determine the structure of the neural network for each division region [Wi, Xj, Yk, Zl] so that the difference between the output values y ₁ , y ₂ , y ₃ and the corresponding training data is reduced in advance Learn the weights of the neural network for each divided region [Wi, Xj, Yk, Zl].

Therefore, in this modification, the previously learned division area [Wi, Xj, Yk, Zl] is also referred to as a learned division area [Wi, Xj, Yk, Zl] hereinafter. It should be noted that the training data obtained in advance by actual measurement of the values of the operating parameters of the plurality of categories related to the air conditioner within the preset ranges Rw, Rx, Ry, and Rz are stored in the memory of the electronic control unit 56 . department. Also in this modification, for each of the divided regions [Wi, Xj, Yk, Zl], a neural network having the same structure as the neural network used in the previous learning is used, and the weights of the neural network at the time of completion of the learning are used, and the vehicle is running. Further learning is carried out in the vehicle mode. FIG. 25 shows a learning processing routine of this modification carried out in the vehicle-mounted manner, and the learning processing routine is executed by interruption at regular intervals (for example, every one second).

Referring to FIG. 25 , first, in step 500, the learned weights stored in the storage unit of the electronic control unit 56 and the training data used in the previous learning, that is, the preset ranges Rw, Rx, Ry, The training data obtained in advance by the actual measurement of the values of the operating parameters of the plurality of categories related to the air conditioner in Rz, and each learned divided area [Wi, Xj, Yk, Zl]. This learned weight is used as an initial value for the weight. Next, in step 501, the number Ki, j, k, l of nodes in the previous hidden layer of the output layer used in the previous learning for each learned partition region [Wi, Xj, Yk, Zl] is read in . Next, go to step 502, obtain new input values x1, x2, x3, x4, namely the temperature, humidity, location and the size of the room where the air conditioner is installed, and the new input values x1, x2, x3, x4 are the new and air conditioners The values of the operation parameters of the related plurality of categories are stored in the storage unit of the electronic control unit 56 . Then, in step 502, the air volume, air direction, and operation time of the air conditioner with respect to the new input values x1, x2, x3, and x4 are stored in the storage unit of the electronic control unit 56 as training data. That is, in step 502 , training data obtained by actual measurement of the newly acquired values of the operating parameters of the plurality of categories related to the air conditioner are stored in the storage unit of the electronic control unit 56 .

Next, in step 503, it is determined whether or not the new input values x1, x2, x3, and x4 are within the learned division area [Wi, Xj, Yk, Z1], that is, the newly acquired operation parameters of a plurality of categories related to the air conditioner Whether the value of is within the preset ranges Rw, Rx, Ry, Rz. When the new input values x1, x2, x3, and x4 are within the learned division regions Rw, Rx, Ry, and Rz, that is, when the newly acquired values of the operating parameters of the plurality of categories related to the air conditioner are within the preset values If it is within the ranges Rw, Rx, Ry, and Rz, the process proceeds to step 504, and the new input values x1, x2, x3, and x4, that is, the values of the newly acquired operating parameters of multiple categories related to the air conditioner, are converted to the newly acquired values related to the air conditioner. The input of each node of the input layer of the neural network of the learned partition region [Wi, Xj, Yk, Zl] to which the values of the operating parameters of the plurality of categories belong, is based on the output value y ₁ output from the node of the output layer of the neural network , y ₂ , y ₃ , and training data obtained by actual measurement for the newly acquired values of the operating parameters of a plurality of categories related to the air conditioner, using the error back propagation method, so that the output values y ₁ , y ₂ , y _3. The weights of the neural network of the learned division area [Wi, Xj, Yk, Zl] to which the newly acquired values of the operating parameters of the plurality of categories related to the internal combustion engine belong are further learned so that the difference from the training data becomes smaller.

On the other hand, when it is determined in step 503 that the new input values x1, x2, x3, and x4 are not within the learned division area [Wi, Xj, Yk, Z1], the process proceeds to step 505. Outside the ranges Rw, Rx, Ry, and Rz, an unlearned area defined by the new input values x1, x2, x3, and x4 is set. For example, when it is determined that the new input values x2, x3, and x4 are within the corresponding preset ranges Rx, Ry, and Rz and the new input value x1 is not within the corresponding preset range Rw, if The range to which the new input value x1 belongs is set to Wa, and the unlearned area [Wa, Xj, Yk, Zl] delimited by the new input values x1, x2, x3, and x4 is set. In addition, when it is determined that the new input values x3 and x4 are within the corresponding preset ranges Ry and Rz and the new input values x1 and x2 are not within the corresponding preset ranges Rw and Rx, if the new input values The range to which the new input value x1 belongs is set to Wa and the range to which the new input value x2 belongs is set to Xb, then the unlearned area delimited by the new input values x1, x2, x3, and x4 is set [Wa, Xb, Yk, Zl].

Next, in step 505, a new neural network for the unlearned region [Xa, Yb] is created. After creating a new neural network in step 505 , go to step 504 . In step 504, with regard to the unlearned area, the weights of the new neural network created for the unlearned area are learned so that the difference between the output values y ₁ , y ₂ , and y ₃ and the training data becomes small.

26 to 33 show a fifth example of a case where the machine learning apparatus of the present invention is applied to estimation of the degree of deterioration of a secondary battery. In this embodiment, the degree of deterioration of the secondary battery is detected based on the air temperature, the temperature of the secondary battery, the discharge time of the secondary battery, and the discharge energy per unit time of the secondary battery. In this case, the range of use conditions and modes of use of the secondary battery, that is, the operating parameters of the secondary battery, such as air temperature, temperature of the secondary battery, discharge time of the secondary battery, and discharge energy per unit time of the secondary battery, etc. The use range of the value can be pre-estimated according to the type of the secondary battery. Therefore, the range of the value of the operating parameter of the secondary battery is usually set in advance so that the output value of the neural network and the actual measured value of the secondary battery are set in advance. The weights of the neural network are learned in advance so that the difference in the degree of deterioration becomes smaller.

However, in this case, the value of the operating parameter of the secondary battery may be outside the preset range. In this case, learning based on the actual value is not performed outside the preset range. Since the output value calculated using the neural network is a value that deviates significantly from the actual value, even in this embodiment, when the newly acquired value of the operating parameter related to the secondary battery is outside the preset range , increase the number of nodes in the previous hidden layer of the output layer of the neural network, or increase the number of neural networks, using the training data obtained from the newly acquired values of the operating parameters related to the secondary battery and the The weights of the neural network are learned from the training data obtained from the values of the operating parameters related to the secondary battery within the preset range.

Next, the fifth embodiment will be specifically described. 26, 60 denotes a secondary battery, 61 denotes an electric motor, 62 denotes a drive control device for the electric motor 61, 63 denotes a voltmeter for detecting the voltage between the output terminals of the secondary battery 60, and 64 denotes a voltmeter for detecting the voltage between the output terminals of the secondary battery 60. The ammeter for the current supplied from the secondary battery 60 to the electric motor 61 via the drive control device 62; The electronic control unit 30 shown is an electronic control unit of the same structure. As shown in FIG. 26 , the supply current to the electric motor 61 detected by the ammeter 53 , the voltage between the output terminals of the secondary battery 60 detected by the voltmeter 64 , the air temperature detected by the thermometer 65 , and the temperature sensor 66 The detected temperature of the secondary battery 60 is input to the electronic control unit 56 , and the electronic control unit 56 calculates an estimated value of the degree of deterioration of the secondary battery 60 . In the electronic control unit 56 , the discharge time of the secondary battery 60 is obtained based on the detection value of the ammeter 64 , and the secondary battery is obtained based on the detection value of the ammeter 64 and the detection value of the voltmeter 63 . Discharge energy (current/voltage) per unit time of 60.

FIG. 27 shows the neural network used in this fifth embodiment. In this fifth embodiment, as shown in FIG. 27 , the input layer (L=1) of the neural network is composed of four nodes, and input value x 1 representing the temperature and an input value x ₁ representing the temperature of the secondary battery 60 are input to each node. The input value x ₂ , the discharge time x ₃ of the secondary battery 60 , and the input value x ₄ representing the discharge energy per unit time of the secondary battery 60 are input. In addition, the number of hidden layers (L=2, L=3) can be set to one or an arbitrary number, and the number of nodes of the hidden layers (L=2, L=3) can also be set to any number number. In addition, in this fourth embodiment, the node of the output layer (L=4) is set to one, and the output value y indicating the degree of deterioration of the secondary battery 20 is output from the node.

_On the other hand, in FIG. 28A , between A1 and B1, that is, R1 represents _a preset range _of air temperature ₍ for example, -5°C to 40°C ₎ , and between A2 and B2, that is, R2 represents _two The pre-set range of the temperature of the secondary battery 60 (for example, -40° C. to ₄₀ ° C.), between A3 and _B3 , that is, _R3 represents the pre _- set range of the discharge time of the secondary battery 60, and A4 and B3 Between B ₄ , that is, R ₄ represents a predetermined range of the discharge energy per unit time of the secondary battery 60 . 28B is also the same as FIG. 28A , between A1 and B1 shows the preset range _of the air temperature, and between _A2 and _B2 shows the preset range _of the temperature of the secondary battery 60, _The interval between A3 and _B3 represents a preset range of the discharge time of the secondary battery 60 , and the interval between _A4 and B4 represents a preset range of the discharge energy per unit time of the secondary battery 60 _.

Here, the relationship between the air temperature, the temperature of the secondary battery 60, the discharge time of the secondary battery 60, the discharge energy per unit time of the secondary battery 60, and the degree of deterioration of the secondary battery 60 will be briefly described. As the secondary battery 60 deteriorates, the internal resistance increases. Therefore, the degree of deterioration of the secondary battery 60 can be estimated from the change in the internal resistance. In practice, however, it is difficult to detect the internal resistance. On the other hand, when the discharge current is constant, the higher the internal resistance, the higher the calorific value of the secondary battery 60. Therefore, the higher the internal resistance, that is, the more the secondary battery 60 is degraded, the temperature of the secondary battery 60 increases. higher. Therefore, the degree of deterioration of the secondary battery 60 can be estimated based on the temperature rise amount of the secondary battery 60 . In this case, the amount of temperature increase of the secondary battery 60 is affected by the air temperature, and is also affected by the discharge time of the secondary battery 60 and the discharge energy per unit time of the secondary battery 60 . Therefore, the degree of deterioration of the secondary battery 60 is obtained from the air temperature, the temperature of the secondary battery 60 , the discharge time of the secondary battery 60 , and the discharge energy per unit time of the secondary battery 60 . The temperature of the secondary battery 60 , the discharge time of the secondary battery 60 , and the discharge energy per unit time of the secondary battery 60 can estimate the degree of deterioration of the secondary battery 60 .

On the other hand, when the secondary battery 60 is degraded, the amount of charge to the secondary battery 60 decreases. In this case, when the circuit of the secondary battery 60 is closed immediately after the charging of the secondary battery 60 is completed, a phenomenon proportional to the amount of charging to the secondary battery 60 occurs between the output terminals of the secondary battery 60 . voltage. That is, the voltage between the output terminals of the secondary battery 60 detected by the voltmeter 64 immediately after the charging of the secondary battery 60 is completed is proportional to the amount of charging of the secondary battery 60 . Therefore, the degree of deterioration of the secondary battery 60 can be detected based on the detection voltage of the voltmeter 64 immediately after the charging of the secondary battery 60 is completed. Therefore, in this fifth embodiment, the degree of deterioration of the secondary battery 60 detected from the detected voltage of the voltmeter 64 immediately after the charging of the secondary battery 60 is completed is used as training data for the output value y.

Next, with reference to FIGS. 29 and 30 , a routine for calculating the discharge time of the secondary battery 60 and the like and a routine for acquiring training data, which are executed in the electronic control unit 67 , will be described. Referring to FIG. 29 showing the calculation routine, in step 600 , the discharge time of the secondary battery 60 is calculated from the output value of the ammeter 64 . Next, in step 601 , the discharge energy per unit time of the secondary battery 60 is calculated from the output value of the ammeter 64 and the output value of the voltmeter 63 .

On the other hand, referring to FIG. 30 showing the training data acquisition processing routine, first, in step 610 , it is determined whether or not the charging process to the secondary battery 60 is being performed. When the charging process to the secondary battery 60 is not performed, the process cycle is completed. On the other hand, when the charging process to the secondary battery 60 is being performed, the process proceeds to step 611 to determine whether or not the charging of the secondary battery 60 has been completed. When it is determined that the charging of the secondary battery 60 has been completed, the process proceeds to step 612, and it is determined whether or not the training data request flag set when the training data is requested is set. The training data request flag will be described later. The processing loop completes when the training data required flag is not set. On the other hand, when the training data request flag is set, the process proceeds to step 613 , and the degree of deterioration of the secondary battery 60 is detected based on the detected voltage of the voltmeter 64 . Next, the process proceeds to step 614, and the additional learning flag is set.

Also in this fifth embodiment, with respect to various input values xn ( _n =1, 2, 3, 4) within a preset range Rn, that is, the air temperature, the temperature of the secondary battery 60, the secondary battery 60 The degree of deterioration of the secondary battery 60 based on the discharge time and the discharge energy per unit time of the secondary battery 60 are obtained in advance as training data. The training data are obtained in advance for the values of the operating parameters of the various categories by actual measurement, and the structure of the neural network is determined based on the values of the operating parameters of the plurality of categories related to the secondary battery 60 and the training data so that the output value y is equal to the value of the training data. The weights of the neural network are learned in advance so that the difference between the training data corresponding to the values of the plurality of types of operating parameters related to the secondary battery 60 becomes smaller. The training data obtained in advance by actual measurement of the values of the plurality of types of operating parameters related to the secondary battery 60 within the preset range Rn are stored in the storage unit of the electronic control unit 67 .

Also in this fifth embodiment, a neural network having the same structure as the neural network used in the previous learning is used, and the weight of the neural network at the time of completion of learning is used, and the learning is further performed on-board while the vehicle is running. FIG. 31 shows a learning processing routine of the fifth embodiment carried out in the vehicle-mounted manner, and the learning processing routine is executed by interruption at regular intervals (for example, every one second).

31 , first, in step 700, the learned weights stored in the storage unit of the electronic control unit 67 and the training data used in the previous learning, that is, the sum of the weights within the preset range Rn, are read. Training data obtained in advance by actual measurement for the values of the operating parameters of the plurality of types related to the secondary battery 60 , and preset values indicating the range of the input data, that is, the values of the operating parameters of the plurality of types related to the secondary battery 60 . Range of values An, Bn (n=1, 2, 3, 4) (FIG. 28A). This learned weight is used as an initial value for the weight. Next, in step 701, the number K of nodes in the hidden layer immediately preceding the output layer of the neural network used in the previous learning is read. Next, in step 702, it is determined whether or not the additional learning flag is set. When the additional learning flag is not set, the process proceeds to step 703 .

In step 703 , a new input value x, that is, a new value of a plurality of types of operation parameters related to the secondary battery 60, is obtained, and the new input value x is a new input value x that is a new value of the plurality of types of the secondary battery 60. The values of the operating parameters are stored in the storage unit of the electronic control unit 67 .

Next, in step 704, it is determined whether or not the new input value xn, that _is , the newly acquired values of the operating parameters of the plurality of types related to the secondary battery 60, is within the preset range Rn (between An and Bn), That is, whether or not the new input value x _n is greater than or equal to An and less than or equal to Bn. When the new input value xn is within the preset range _Rn , the process proceeds to step 705, and each input value xn, that _is , the newly acquired values of the operation parameters of the plurality of types related to the secondary battery 60, are sent to the neural network. The corresponding node input of the input layer is based on the output value y output from the node of the output layer of the neural network and the training data obtained by the actual measurement for the newly acquired values of the operation parameters of the plurality of categories related to the secondary battery 60 . , using the error back-propagation method to learn the weights of the neural network in such a way that the difference between the output value y and the training data becomes smaller.

On the other hand, it is determined in step 704 that the new input value x _n , that is, the value of at least one type of the operation parameter value of the plurality of types of operation parameter values related to the secondary battery 60 newly acquired is not within the preset value In the range Rn (between An and Bn) of , for example, in FIG. 28B , the input value x ₁ indicating that the air temperature is within a predetermined range (B ₁ to C 1 ) of B ₁ to C ₁ (B ₁ <C ₁ ) ₁ ), or in FIG. 28B , the input value x ₃ indicating the discharge time of the secondary battery 60 is within a preset range (C ₃ to A ₃ ) of C ₃ to A ₃ (C ₃ <A ₃ ). In the case of , go to step 706 . In step 706, a training data request flag is set, and the new input value xn obtained at this time _is stored as a new input value _xn for additional learning. Next, the processing loop is completed.

When the training data request flag is set, it can be seen from the training data acquisition processing routine of FIG. 30 that when the charging of the secondary battery 60 is completed, the degree of deterioration of the secondary battery 60 is detected, and the degree of deterioration of the secondary battery 60 is used as a The training data for additional learning is stored. Next, the additional learning flag is set. When the additional learning flag is set in step 706 , the process proceeds from step 702 to step 707 in the next processing loop. In step 707, the new input value x _n stored for additional learning and the training data stored for additional learning are read from the storage unit, and the range (B _n ) to which the new input value x _n belongs is calculated. ～C _n ) or the density D of the training data in the range (C _n ～A _n ) relative to the new input value x _n (= number of training data/(C _n -B _n ) or number of training data/(A _n -C _n )). The definition of the training data density D is as described above. In step 707, after the training data density D is calculated, it is determined whether the training data density D becomes higher than the predetermined data density D ₀ . In the event that the training data density D is lower than the predetermined data density D ₀ , the processing loop is completed.

On the other hand, when it is determined in step 707 that the training data density D has become higher than the predetermined data density D ₀ , the process proceeds to step 708 . In this case, when D(=number of training data/(A _n -C _n ))>D ₀ , the number of additional nodes α is calculated by the following equation.

Number of additional nodes α=round{(K/(Bn-An))·(An-Cn)}

On the other hand, when D (=number of training data/(C _n -B _n ))>D ₀ , the number of additional nodes α is calculated by the following equation.

Number of additional nodes α=round{(K/(Bn-An))·(Cn-Bn)}

It should be noted that, in the above formula, K represents the number of nodes, and round means rounding.

After calculating the number of additional nodes α in step 708, the process proceeds to step 709, where the number K of nodes in the previous hidden layer of the output layer of the neural network is updated to increase the number K of nodes in the previous hidden layer of the output layer The number of nodes α (K←K+α) is added. In this way, in the fifth embodiment, when the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value of the preset range representing the value of the operating parameter increases, the data density increases. The number of nodes in the previous hidden layer of the output layer of the neural network. That is, in the fifth embodiment, the data density is increased in accordance with the increase in the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value of the preset range representing the value of the operating parameter. Increase the number of nodes in the previous hidden layer of the output layer of the neural network.

When the number K of nodes in the previous hidden layer of the output layer is increased by the number of additional nodes α in step 709 (K←K+α), the process proceeds to step 710, so that the number of nodes in the previous hidden layer of the output layer is increased. The neural network is updated by increasing the number K. Next, go to step 705 . In step 705, the training data newly obtained with respect to the new input value x is also included in the training data, and the weight of the updated neural network is learned so that the difference between the output value y and the training data becomes smaller. That is, in step 705 , the training data obtained by actual measurement of the newly acquired values of the operation parameters of the plurality of types related to the secondary battery 60 and the data related to the secondary battery 60 within the preset range Rn are used. The training data obtained in advance by the actual measurement of the values of the operating parameters of the plurality of types of the The weights of the updated neural network are learned so that the difference between the output value y which changes in value and the training data corresponding to the values of the operating parameters of the plurality of categories related to the secondary battery 60 becomes smaller.

32 and 33 show a modification of the fifth embodiment. In this modification, the predetermined range of the value of the operating parameter of each type related to the secondary battery 60 is divided into a plurality of groups. That is, Rw, Rx, Ry, and Rz in FIG. 32 represent the predetermined ranges of the air temperature, the temperature of the secondary battery 60, the discharge time of the secondary battery 60, and the discharge energy per unit time of the secondary battery 60, respectively. As shown in FIG. 24 , the predetermined ranges of the air temperature, the temperature of the secondary battery 60 , the discharge time of the secondary battery 60 , and the discharge energy per unit time of the secondary battery 60 are divided into a plurality of ranges. It should be noted that, in FIG. 32 , W ₁ , W ₂ . . . W _n , X ₁ , X ₂ . . . _{X n} , _{Y 1} , _{Y 2 .} The division range of the value of the operating parameter.

Furthermore, in this modification, a plurality of divided regions [Wi, Xj, Yk, Zl] (i=1) demarcated by the combination of the divided ranges obtained by dividing the values of the operating parameters of each category are preset. , 2...n, j=1, 2...n, k=1, 2...n, l=1, 2...n), an independent neural network is created for each divided area [Wi, Xj, Yk, Zl] . These neural networks have the configuration shown in FIG. 27 . In this case, the number of nodes in the hidden layer (L=3) is different for each neural network. Hereinafter, the hidden layer before the output layer of the neural network in the region [Wi, Xj, Yk, Zl] will be divided The number of nodes is represented by Ki, j, k, l. The numbers Ki, j, k, and l of nodes in the hidden layer are set in advance according to the complexity of the change of the training data in each of the divided regions [Wi, Xj, Yk, Zl] with respect to the change of the input value.

In this modified example, each of the divided regions [Wi, Xj, Yk, Zl formed in the preset ranges Rw, Rx, Ry, Rz of the values of the operating parameters of the plurality of types related to the secondary battery 60 The various input values x1 , x2 , x3 , and x4 in the The degree of deterioration of , is obtained in advance as training data, that is, the values of a plurality of types of operating parameters related to the secondary battery 60 within the preset ranges Rw, Rx, Ry, and Rz are obtained in advance by actual measurement. The training data is determined for each divided area [Wi, The structure of the neural network of Xj, Yk, Zl] learns in advance the difference between the output values y ₁ , y ₂ , y ₃ and the corresponding training data to reduce the difference between the divided regions [Wi, Xj, Yk, Zl] Neural network weights.

Therefore, in this modification, the previously learned division area [Wi, Xj, Yk, Zl] may also be referred to as a learned division area [Wi, Xj, Yk, Zl] hereinafter. It should be noted that training data obtained in advance by actual measurement of the values of the plurality of types of operating parameters related to the secondary battery 60 within the preset ranges Rw, Rx, Ry, and Rz are stored in the electronic control unit. 56 storage section. Also in this modification, for each of the divided regions [Wi, Xj, Yk, Zl], a neural network having the same structure as the neural network used in the previous learning is used, and the weights of the neural network at the time of completion of the learning are used, and the vehicle is running. Further learning is carried out in the vehicle mode. FIG. 33 shows a learning processing routine of this modification carried out in the vehicle-mounted manner, and the learning processing routine is executed by interruption at regular intervals (for example, every one second).

Referring to FIG. 33 , first, in step 800, the learned weights stored in the storage unit of the electronic control unit 56 and the training data used in the previous learning, that is, the preset ranges Rw, Rx, Ry, The training data obtained in advance by actual measurement of the values of the operating parameters of the plurality of types related to the secondary battery 60 in Rz, and each learned division area [Wi, Xj, Yk, Zl]. This learned weight is used as an initial value for the weight. Next, in step 801, the number Ki, j, k, l of nodes in the previous hidden layer of the output layer used in the previous learning for each learned partition region [Wi, Xj, Yk, Zl] is read in . Next, in step 802, it is determined whether or not the additional learning flag is set. When the additional learning flag is not set, the process proceeds to step 803 .

In step 803, new input values x1, x2, x3, and x4 are obtained, namely the air temperature, the temperature of the secondary battery 60, the discharge time of the secondary battery 60, and the discharge energy per unit time of the secondary battery 60. The new input values x1, x2, x3, and x4 are obtained. The input values x1 , x2 , x3 , and x4 , that is, the new values of the operation parameters of a plurality of types related to the secondary battery 60 are stored in the storage unit of the electronic control unit 56 .

Next, in step 804, it is determined whether the new input values x1, x2, x3, and x4 are within the learned division area [Wi, Xj, Yk, Z1], that is, the newly acquired multiple categories related to the secondary battery 60 Whether the value of the operating parameter is within the preset ranges Rw, Rx, Ry, Rz. When the new input values x1 , x2 , x3 , and x4 are within the learned division regions Rw, Rx, Ry, and Rz, that is, when the newly acquired values of the operating parameters of the plurality of categories related to the secondary battery 60 are in the preliminarily If it is within the set ranges Rw, Rx, Ry, and Rz, the process proceeds to step 805, and the input values x1, x2, x3, and x4, that is, the newly acquired values of the plurality of types of operating parameters related to the secondary battery 60, are added to the newly acquired The input of each node of the input layer of the neural network of the learned partition region [Wi, Xj, Yk, Zl] to which the values of the operating parameters of the plurality of categories related to the secondary battery 60 belong, is based on the input from the output layer of the neural network. The output value y output by the node and the training data obtained by actual measurement of the newly acquired values of the operation parameters of the plurality of categories related to the secondary battery 60 are obtained by using the error back propagation method, so that the output value y is related to the training data. The weights of the neural network of the learned division regions [Wi, Xj, Yk, Zl] to which the newly acquired values of the operating parameters of the plurality of types related to the secondary battery 60 belong so that the difference becomes smaller are further learned.

On the other hand, when it is determined in step 804 that the new input values x1, x2, x3, and x4 are not within the learned division area [Wi, Xj, Yk, Z1], the process proceeds to step 806 . In step 806, a training data request flag is set, and the new input value xn acquired at this time _is stored as a new input value _xn for additional learning. Next, the processing loop is completed.

When the training data request flag is set, it can be seen from the training data acquisition processing routine of FIG. 30 that when the charging of the secondary battery 60 is completed, the degree of deterioration of the secondary battery 60 is detected, and the degree of deterioration of the secondary battery 60 is used as a The training data for additional learning is stored. Next, the additional learning flag is set. After the additional learning flag is set in step 806 , the process proceeds from step 802 to step 807 in the next processing loop. In step 807, the new input value _xn stored for additional learning and the training data stored for additional learning are read from the storage unit, and are outside the preset ranges Rw, Rx, Ry, Rz An unlearned area defined by the new input values x1, x2, x3, and x4 stored for additional learning is set. For example, when it is determined that the new input values x2, x3, and x4 are within the corresponding preset ranges Rx, Ry, and Rz and the new input value x1 is not within the corresponding preset range Rw, if the The range to which the new input value x1 belongs is set to Wa, and an unlearned area [Wa, Xj, Yk, Zl] defined by the new input values x1, x2, x3, and x4 is set. In addition, when it is determined that the new input values x3 and x4 are within the corresponding preset ranges Ry and Rz and the new input values x1 and x2 are not within the corresponding preset ranges Rw and Rx, if the new input values The range to which the new input value x1 belongs is set to Wa, and the range to which the new input value x2 belongs is set to Xb, then the unlearned area delimited by the new input values x1, x2, x3, and x4 is set [Wa, Xb, Yk, Zl].

Next, in step 807, a new neural network for the unlearned region is created. After creating a new neural network in step 807 , go to step 805 . In step 805 , the weights of the new neural network created for the unlearned area are learned so that the difference between the output value y and the training data stored for additional learning becomes small about the unlearned area.

From the above description, in the embodiment of the present invention, in the machine learning device for outputting the output value with respect to the value of the operating parameter of the machine using the neural network, the specific type of the machine is preset. The value range of the operating parameter is preset, and the number of nodes of the hidden layer of the neural network corresponding to the value range of the specific type of operating parameter related to the device is preset. When the newly acquired value of the operation parameter of the specific type related to the machine is outside the preset range, increase the number of nodes in the previous hidden layer of the output layer of the neural network, and use the newly acquired and The weights of the neural network are learned from training data obtained by actual measurement of the values of the specific types of operating parameters related to the equipment and training data obtained by actual measurement of the values of the operating parameters of the equipment within a preset range. A neural network that has learned weights is used to output an output value relative to the value of a particular class of operating parameters associated with the machine.

In this case, in the embodiment of the present invention, the machine learning device includes an electronic control unit. The electronic control unit includes: a parameter value acquisition unit that acquires values of operating parameters of a specific type related to the above-mentioned equipment; a calculation unit that performs calculations using a neural network including an input layer, a hidden layer, and an output layer; and a storage unit, The value of the operating parameter of the specific type related to the above-mentioned equipment is input to the input layer, and the output value that varies according to the value of the specific type of operating parameter related to the above-mentioned equipment is output from the output layer. The range of values of the operating parameters of the specific type related to the above-mentioned equipment is preset, and the nodes of the hidden layer of the neural network corresponding to the value range of the specific type of operating parameters related to the above-mentioned equipment are preset. The number of objects is stored in the storage unit as training data obtained in advance by actual measurement for the values of the operating parameters of the specific type related to the above-mentioned equipment within a preset range. When the newly acquired value of the operating parameter of the specific type related to the above-mentioned equipment is within a preset range, a training method obtained by actual measurement of the newly acquired value of the specific type of operating parameter related to the above-mentioned equipment is used The data is used to make the output value changed according to the newly acquired value of the operating parameter of the specific type related to the above-mentioned equipment and the value of the newly acquired operating parameter of the specific type related to the above-mentioned equipment to pass actual measurement by using the calculation unit The weights of the neural network are learned in such a way that the difference between the obtained training data becomes smaller. When the value of the operation parameter of the specific type related to the above-mentioned equipment newly acquired by the parameter value acquisition unit is outside the preset range, the value of the newly acquired operation parameter of the specific type related to the above-mentioned equipment is passed through The increase in the number of training data obtained by the actual measurement or the increase in the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value of the preset range representing the value of the operating parameter corresponds to an increase in the data density. In order to increase the number of nodes in the previous hidden layer of the output layer of the neural network, and use the training data obtained by the actual measurement of the newly acquired values of the operating parameters of the specific category related to the above-mentioned equipment, and obtained in advance The training data is used to make the output values that vary according to the values of the specific types of operating parameters related to the above-mentioned equipment within the preset range and outside the preset range to correspond to the above-mentioned equipment. The weights of the neural network are learned so that the difference between the training data of the values of the operation parameters of the specific category becomes smaller. An output value with respect to a value of a specific class of operating parameters related to the above-mentioned machine is output using a neural network that has learned weights.

In addition, in this case, in the embodiment of the present invention, the electronic control unit includes: a parameter value acquisition unit that acquires the value of the operation parameter of the specific type related to the above-mentioned equipment; and a calculation unit that uses the input layer and the hidden layer and the neural network of the output layer to perform operations; and a storage unit, the value of the specific type of operating parameter related to the above-mentioned equipment is input to the input layer, and the value of the specific type of operating parameter related to the above-mentioned equipment is changed. Multiple output values are output from the output layer. The range of values of the operating parameters of the specific type related to the above-mentioned equipment is preset, and the nodes of the hidden layer of the neural network corresponding to the value range of the specific type of operating parameters related to the above-mentioned equipment are preset. The number of objects is stored in the storage unit as training data obtained in advance by actual measurement for the values of the operating parameters of the specific type related to the above-mentioned equipment within a preset range. When the value of the operating parameter of the above-mentioned equipment newly acquired by the parameter value acquisition unit is within the preset range, the training obtained by the actual measurement of the newly acquired value of the operating parameter of the specific type related to the above-mentioned equipment is used The data is obtained by using the arithmetic unit to obtain a difference between a plurality of output values that vary according to the value of the specific type of operating parameter related to the above-mentioned equipment and training data corresponding to the value of the specific type of operating parameter related to the above-mentioned equipment Learning the weights of a neural network in a smaller way. When the value of the operation parameter of the specific type related to the above-mentioned equipment newly acquired by the parameter value acquisition unit is outside the preset range, the value of the operation parameter of the specific type related to the above-mentioned equipment newly acquired is passed through The increase of the number of training data obtained by actual measurement or the increase of the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value representing a preset range increases the performance of the neural network. The number of nodes in the previous hidden layer of the output layer is increased, and the values of the operating parameters of the specific category related to the above-mentioned equipment that are newly acquired within the preset range and outside the preset range are used to pass the actual measurement. The obtained training data and the previously obtained training data are used by the arithmetic unit to make a plurality of output values that vary according to the value of the operating parameter of the specific type related to the above-mentioned equipment to correspond to the specific type of the above-mentioned equipment. The weights of the neural network are learned so that the difference between the values of the operating parameters of the classes and the training data becomes smaller. A plurality of output values with respect to the values of the above-described specific types of operating parameters related to the machine are output using the neural network that has learned the weights.

On the other hand, in the embodiment of the present invention, in the machine learning device for outputting the output value with respect to the value of the operating parameter of the machine using the neural network, a plurality of categories related to the machine are preset. The value range of the operating parameter, and the number of nodes in the hidden layer of the neural network corresponding to the range of values of the operating parameters of a plurality of categories related to the machine is preset, and the newly acquired data related to the machine is used. When the values of the operating parameters of multiple categories are outside the preset range, the number of nodes in the previous hidden layer of the output layer of the neural network is increased, and the newly acquired values of multiple categories related to the machine are used. The weights of the neural network are learned from the training data obtained by the actual measurement of the values of the operation parameters and the training data obtained by the actual measurement of the values of the operation parameters of the equipment within a preset range. The neural network that has learned the weights is used to output output values with respect to the values of the operating parameters of the plurality of categories related to the machine.

In this case, in the embodiment of the present invention, the machine learning device includes an electronic control unit including: a parameter value acquisition unit that acquires values of a plurality of types of operating parameters related to the above-mentioned equipment; a calculation unit , using a neural network including an input layer, a hidden layer, and an output layer to perform operations; and a storage unit, in which the values of a plurality of types of operating parameters related to the above-mentioned equipment are input to the input layer, Output values that vary depending on the values of the operating parameters of each category are output from the output layer. For each of the plurality of types of operating parameters related to the above-mentioned equipment, a range of values of the operating parameters of each type is preset, and a value of the plurality of types of operating parameters related to the above-mentioned equipment is preset. The number of nodes in the hidden layer of the neural network corresponding to the range, the values of the operating parameters of a plurality of categories are obtained in advance through actual measurement, and the training data of the values of the operating parameters of each category within the preset range is stored in storage department. When the values of the plurality of operating parameters of the above-mentioned equipment newly acquired by the parameter value acquisition unit are within the preset ranges, respectively, the actual measurement is performed using the newly acquired values of the operating parameters of the plurality of types related to the above-mentioned equipment. The training data thus obtained is used by the computing unit to make the output values that vary according to the values of the operating parameters of the plurality of categories related to the above-mentioned equipment and the values of the operating parameters corresponding to the plurality of categories related to the above-mentioned equipment to be The weights of the neural network are learned in such a way that the difference between the training data becomes smaller. When the value of the operation parameter of at least one of the plurality of types of operation parameters related to the above-mentioned equipment newly acquired by the parameter value acquisition unit is outside the preset range, the newly acquired operation parameter related to the above-mentioned equipment is The values of the operating parameters of a plurality of categories are obtained by increasing the number of training data obtained by actual measurement or by dividing the number of training data by the difference between the maximum value and the minimum value representing a preset range. The increase of the density correspondingly increases the number of nodes in the previous hidden layer of the output layer of the neural network, and uses the newly obtained data within the preset range and outside the preset range with the above-mentioned The training data obtained by the actual measurement of the values of the operating parameters of the plurality of categories related to the equipment and the training data obtained in advance are used by the computing unit to change according to the values of the operating parameters of the plurality of categories related to the above-mentioned equipment. The weights of the neural network are learned such that the difference between the output value and the training data corresponding to the values of the operating parameters of the above-described plurality of categories of equipment becomes smaller. Output values for the values of the operating parameters of the plurality of categories related to the above-mentioned equipment are output using the neural network that has learned the weights.

In addition, in this case, in the embodiment of the present invention, the electronic control unit includes: a parameter value acquisition unit that acquires the values of a plurality of types of operation parameters related to the above-mentioned equipment; layer and the output layer of the neural network to perform operations; and a storage unit, the values of the plurality of types of operating parameters related to the above-mentioned equipment are input to the input layer, and the values of the plurality of types of operating parameters related to the above-mentioned equipment are used. And the changed multiple output values are output from the output layer. For each of the plurality of types of operating parameters related to the above-mentioned equipment, a range of values of the operating parameters of each type is preset, and a value of the plurality of types of operating parameters related to the above-mentioned equipment is preset. The number of nodes in the hidden layer of the neural network corresponding to the range, the values of the operating parameters of a plurality of categories are obtained in advance through actual measurement, and the training data of the values of the operating parameters of each category within the preset range is stored in storage department. When the values of the plurality of operating parameters of the above-mentioned equipment newly acquired by the parameter value acquisition unit are within the preset ranges, respectively, the actual measurement is performed using the newly acquired values of the operating parameters of the plurality of types related to the above-mentioned equipment. The training data thus obtained is used by the arithmetic unit so that a plurality of output values that vary according to the values of the plurality of types of operating parameters related to the above-mentioned equipment and the plurality of types of operating parameters corresponding to the above-mentioned equipment are obtained. The weights of the neural network are learned in such a way that the difference between the values of the training data becomes smaller. When the value of the operation parameter of at least one of the plurality of types of operation parameters related to the above-mentioned equipment newly acquired by the parameter value acquisition unit is outside the preset range, the newly acquired operation parameter related to the above-mentioned equipment is The values of the operating parameters of a plurality of categories are obtained by increasing the number of training data obtained by actual measurement or by dividing the number of training data by the difference between the maximum value and the minimum value representing a preset range. The increase of the density correspondingly increases the number of nodes in the previous hidden layer of the output layer of the neural network, and uses the newly acquired and the above-mentioned machines within the preset range and outside the preset range. The training data obtained by the actual measurement and the training data obtained in advance for the values of the operating parameters of the plurality of categories related to the above-mentioned apparatuses are used to change more according to the values of the operating parameters of the plurality of categories related to the above-mentioned equipment. The weights of the neural network are learned such that the difference between each output value and the training data corresponding to the values of the operating parameters of the plurality of categories related to the above-mentioned equipment becomes smaller. A plurality of output values for the values of the operating parameters of the plurality of categories related to the above-mentioned equipment are output using the neural network that has learned the weights.

On the other hand, in the embodiment of the present invention, in the machine learning device for outputting the output value with respect to the value of the operating parameter of the machine using the neural network, a plurality of categories related to the machine are preset. The range of values of the operating parameters, and the neural network corresponding to the range of values of the operating parameters of the plurality of categories related to the device is pre-formed, among the newly acquired values of the operating parameters of the plurality of categories related to the device When the value of the operating parameter of at least one of the categories is outside the preset range, a new neural network is formed, and the training data obtained by the actual measurement of the newly acquired values of the operating parameters of the plurality of categories related to the machine is used to form a new neural network. Learn the weights of a new neural network. The neural network that has learned the weights is used to output output values with respect to the values of the operating parameters of the plurality of categories related to the machine.

In this case, in the embodiment of the present invention, the machine learning device includes an electronic control unit including: a parameter value acquisition unit that acquires values of a plurality of types of operating parameters related to the above-mentioned equipment; a calculation unit , using a plurality of neural networks including an input layer, a hidden layer and an output layer to perform operations; and a storage unit. The values of the operating parameters of the plurality of categories related to the above-mentioned equipment are input to the input layer, and output values that vary according to the values of the plurality of categories of operating parameters related to the above-mentioned equipment are output from the corresponding output layer. For each of the plurality of types of operation parameters related to the above-mentioned equipment, a range of values of the operation parameters of each type is preset, and the preset ranges of the values of the operation parameters of each type are divided into a plurality of groups, and A plurality of divided regions demarcated by the combination of the divided ranges of the values of the operating parameters of each category are set in advance. A neural network is created for each divided region, and the number of nodes of the hidden layer of each neural network is preset. The training data obtained in advance by actual measurement for the values of the operating parameters of the plurality of categories are stored in the storage unit. When the values of the operating parameters of the plurality of categories related to the above-mentioned equipment newly acquired by the parameter value acquisition unit are within a preset range, the values of the newly acquired operating parameters of the plurality of categories related to the above-mentioned equipment are used. The training data obtained by the actual measurement is used to make the output values that vary according to the values of the operating parameters of the plurality of categories related to the above-mentioned equipment and the operating parameters corresponding to the plurality of categories of the above-mentioned equipment by the computing unit. The weights of the neural network of the divided regions to which the values of the operating parameters of the plurality of categories newly acquired related to the above-mentioned equipment belong are learned so that the difference between the values of the training data becomes smaller. When the value of the operation parameter of at least one type of the operation parameters of the plurality of types newly acquired by the parameter value acquisition unit is outside the preset range, the value of the operation parameter of at least one type is set It belongs to a new region defined by a combination of preset ranges of operating parameter values for each category, and a new neural network is created for the new region. Using the training data obtained by the actual measurement of the newly acquired values of the operating parameters of the plurality of categories related to the above-mentioned equipment, the arithmetic unit is used to change the values of the operating parameters of the plurality of categories related to the above-mentioned equipment. The weights of the new neural network are learned in such a way that the difference between the output value of and the training data corresponding to the values of the operating parameters of the plurality of categories related to the above-mentioned machine becomes smaller. Output values corresponding to the values of the above-mentioned operating parameters of the equipment are output using each neural network that has learned the weights.

Label description

1 Internal combustion engine

14 Throttle valve opening sensor

23 NO _X sensor

24 Atmospheric temperature sensor

30, 56, 67 Electronic control unit

50 Main body of air conditioner

53, 65 Thermometer

54 Hygrometer

55 GPS

60 Secondary batteries

53 Galvanometer

64 Voltmeter.

Claims (9)

1. A machine learning device for an internal combustion engine, comprising an electronic control unit, The electronic control unit has: a parameter value acquisition unit for acquiring values of specific types of operating parameters related to the internal combustion engine; an operation part, which uses a neural network including an input layer, a hidden layer and an output layer to perform operations; and storage department, The value of the specific type of operating parameter related to the internal combustion engine is input to the input layer, and the output value that varies according to the value of the specific type of operating parameter related to the internal combustion engine is output from the output layer, In this machine learning device, The range of values of the specific type of operating parameter related to the internal combustion engine is preset, and the number of nodes of the hidden layer of the neural network corresponding to the range of the value of the specific type of operating parameter related to the internal combustion engine is preset. number, The training data obtained in advance by actual measurement of the values of the specific types of operating parameters related to the internal combustion engine within the preset range is stored in the storage unit, When the newly acquired value of the operating parameter of the specific type related to the internal combustion engine is within a preset range, training data obtained by actual measurement of the newly acquired value of the specific type of operating parameter related to the internal combustion engine is used, Using this computing unit, the output value that changes according to the newly acquired value of the operating parameter of the specific type related to the internal combustion engine and the newly acquired value of the operating parameter of the specific type related to the internal combustion engine are obtained by actual measurement. The weights of the neural network are learned in such a way that the difference between the training data becomes smaller, When the value of the operating parameter of the specific type related to the internal combustion engine newly acquired by the parameter value acquisition unit is outside the preset range, the value of the newly acquired operating parameter of the specific type related to the internal combustion engine is measured by actual measurement The increase in the number of training data obtained or the increase in the data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value of a preset range representing the value of the operating parameter corresponds to an increase in the data density. Specifically, the number of nodes in the previous hidden layer of the output layer of the neural network is increased, and the training data obtained by actual measurement of the newly acquired values of the operating parameters of the specific type related to the above-mentioned internal combustion engine and the pre-required values are used. The obtained training data is used to make the output value that changes according to the value of the operating parameter of the specific type related to the above-mentioned internal combustion engine within the preset range and outside the preset range to correspond to the The weights of the neural network are learned such that the difference between the training data of the values of the specific types of operating parameters related to the internal combustion engine becomes smaller, and the weights of the neural network are used to output the values of the specific types of operating parameters related to the internal combustion engine. output value. 2. A machine learning device for an internal combustion engine, comprising an electronic control unit, The electronic control unit has: a parameter value acquisition unit for acquiring values of specific types of operating parameters related to the internal combustion engine; an operation part, which uses a neural network including an input layer, a hidden layer and an output layer to perform operations; and storage department, The value of the specific type of operating parameter related to the internal combustion engine is input to the input layer, and a plurality of output values that vary according to the value of the specific type of operating parameter related to the internal combustion engine are output from the output layer, In this machine learning device, The range of values of the specific type of operating parameter related to the internal combustion engine is preset, and the number of nodes of the hidden layer of the neural network corresponding to the range of the value of the specific type of operating parameter related to the internal combustion engine is preset. number, The training data obtained in advance by actual measurement of the values of the specific types of operating parameters related to the internal combustion engine within the preset range is stored in the storage unit, When the value of the operating parameter of the internal combustion engine newly acquired by the parameter value acquisition unit is within a preset range, training data obtained by actual measurement of the newly acquired value of the operating parameter of the specific type related to the internal combustion engine is used and the difference between the plurality of output values that vary according to the value of the specific type of operating parameter related to the internal combustion engine and the training data corresponding to the value of the specific type of operating parameter related to the internal combustion engine is reduced by the computing unit way to learn the weights of the neural network, When the value of the operating parameter of the specific type related to the internal combustion engine newly acquired by the parameter value acquisition unit is outside the preset range, the value of the newly acquired operating parameter of the specific type related to the internal combustion engine is measured by actual measurement The increase in the number of training data obtained or the increase in data density obtained by dividing the number of training data by the difference between the maximum value and the minimum value representing a preset range corresponds to the increase in the neural network. The number of nodes in the previous hidden layer of the output layer increases, and the newly acquired values of the operating parameters of the specific category related to the internal combustion engine within the preset range and outside the preset range are used. The training data obtained by the actual measurement and the training data obtained in advance are used by the computing unit to make a plurality of output values that vary according to the values of the specific types of operating parameters related to the internal combustion engine to correspond to those related to the internal combustion engine. The weights of the neural network are learned so that the difference between the values of the operating parameters of the specific category of the training data becomes smaller, and the neural network that has learned the weights is used to output a plurality of outputs with respect to the values of the operating parameters of the specific category related to the above-mentioned internal combustion engine. value. 3. A machine learning device for an internal combustion engine or a secondary battery, comprising an electronic control unit, The electronic control unit has: a parameter value acquisition unit for acquiring values of a plurality of types of operating parameters related to the internal combustion engine or the secondary battery; an operation part, which uses a neural network including an input layer, a hidden layer and an output layer to perform operations; and storage department, The values of the plurality of types of operating parameters related to the internal combustion engine or the secondary battery are input to the input layer, and output values that vary according to the values of the plurality of types of operating parameters related to the internal combustion engine or the secondary battery are sent from the output layer. output, In this machine learning device, For each of the plurality of categories of operating parameters related to the above-mentioned internal combustion engine or secondary battery, a range of values of the operating parameters of each category is preset, and a plurality of categories related to the above-mentioned internal combustion engine or secondary battery are preset. The number of nodes in the hidden layer of the neural network corresponding to the range of values of the operating parameters, The values of the operating parameters of a plurality of categories are obtained in advance through actual measurement, and the training data in which the values of the operating parameters of each category are within the preset range are stored in the storage unit, When the values of the plurality of operating parameters of the internal combustion engine or the secondary battery newly acquired by the parameter value acquisition unit are within a preset range, respectively, the plurality of categories related to the internal combustion engine or the secondary battery newly acquired are used. The training data obtained by the actual measurement of the values of the operating parameters, and the computing unit is used to make the output values that vary according to the values of the operating parameters of the plurality of categories related to the internal combustion engine or the secondary battery to correspond to the internal combustion engine or the secondary battery. The weights of the neural network are learned so that the difference between the training data of the values of the operation parameters of the plurality of categories related to the secondary battery becomes smaller, When the value of the operating parameter of at least one of the plurality of types of operating parameters related to the internal combustion engine or the secondary battery newly acquired by the parameter value acquisition unit is outside the preset range, it is compared with the newly acquired and Increase in the number of training data obtained by actual measurement of the values of a plurality of types of operating parameters related to the above-mentioned internal combustion engine or secondary battery, or dividing the number of training data by the maximum value and the minimum value indicating a preset range The increase of the data density obtained by the difference of the values correspondingly increases the number of nodes in the previous hidden layer of the output layer of the neural network, and uses the preset range and the preset value. The training data and the previously obtained training data obtained by the actual measurement of the newly acquired values of the plurality of types of operating parameters related to the internal combustion engine or the secondary battery, which are outside the range, are used by the computing unit so as to be based on the internal combustion engine and the above-mentioned internal combustion engine. The neural network is learned in such a way that the difference between the output value that varies depending on the value of the operating parameter of the plurality of categories related to the internal combustion engine or the secondary battery and the training data corresponding to the value of the operating parameter of the plurality of categories related to the internal combustion engine or the secondary battery becomes smaller. As the weight of the network, an output value corresponding to the value of a plurality of types of operating parameters related to the above-mentioned internal combustion engine or a secondary battery is output using a neural network that has learned the weight. 4. A machine learning device for an internal combustion engine or an air conditioner, comprising an electronic control unit, The electronic control unit has: a parameter value acquisition unit for acquiring values of a plurality of types of operating parameters related to the internal combustion engine or the air conditioner; an operation part, which uses a neural network including an input layer, a hidden layer and an output layer to perform operations; and storage department, The values of the plurality of types of operating parameters related to the internal combustion engine or the air conditioner are input to the input layer, and the plurality of output values that vary according to the values of the plurality of types of operating parameters related to the internal combustion engine or the air conditioner are output from the output layer, In this machine learning device, For each of the plurality of types of operating parameters related to the internal combustion engine or the air conditioner, a range of values of the operating parameters of each type is preset, and a plurality of types of operating parameters related to the internal combustion engine or the air conditioner are preset. The number of nodes in the hidden layer of the neural network corresponding to the range of values, The values of the operating parameters of a plurality of categories are obtained in advance through actual measurement, and the training data in which the values of the operating parameters of each category are within the preset range are stored in the storage unit, When the values of the plurality of operating parameters of the internal combustion engine or the air conditioner newly acquired by the parameter value acquisition unit are within the preset ranges, respectively, the values of the newly acquired operating parameters of the plurality of categories related to the internal combustion engine or the air conditioner are used. Values are training data obtained by actual measurement, and the arithmetic unit is used to make a plurality of output values that vary according to the values of the plurality of types of operating parameters related to the internal combustion engine or the air conditioner to correspond to the plurality of output values related to the internal combustion engine or the air conditioner. The weights of the neural network are learned in such a way that the difference between the values of the operating parameters of each category of the training data becomes smaller, When the value of the operation parameter of at least one of the plurality of types of operation parameters related to the internal combustion engine or the air conditioner newly acquired by the parameter value acquisition unit is out of the preset range, the value of the operation parameter newly acquired with the internal combustion engine or an increase in the number of training data obtained by actual measurement of the values of operating parameters of multiple categories related to the air conditioner, or the difference between the maximum value and the minimum value representing a preset range divided by the number of training data The resulting increase in data density correspondingly increases the number of nodes in the previous hidden layer of the output layer of the neural network, and uses new data within the preset range and outside the preset range. The obtained training data and the previously obtained training data obtained by the actual measurement of the values of the plurality of types of operating parameters related to the internal combustion engine or the air conditioner are used by the computing unit to The weights of the neural network are learned so that the difference between the plurality of output values that vary depending on the values of the operating parameters of the class and the training data corresponding to the values of the operating parameters of the plurality of classes related to the internal combustion engine or the air conditioner becomes smaller, and the learned weights are used. A neural network is used to output a plurality of output values relative to the values of the plurality of categories of operating parameters related to the above-mentioned internal combustion engine or air conditioner. 5. A machine learning device for an internal combustion engine, comprising an electronic control unit, The electronic control unit has: a parameter value acquisition unit for acquiring values of a plurality of types of operating parameters related to the internal combustion engine; an operation part, using a plurality of neural networks including an input layer, a hidden layer and an output layer to perform operations; and storage department, The values of the operating parameters of the plurality of categories related to the internal combustion engine are input to the input layer, and output values that vary according to the values of the operating parameters of the plurality of categories related to the internal combustion engine are output from the corresponding output layer, In this machine learning device, For each of the plurality of types of operating parameters related to the above-mentioned internal combustion engine, a range of values of the operating parameters of each type is preset, and the preset range of values of the operating parameters of each type is divided into a plurality of groups, and A plurality of divided regions demarcated by the combination of the divided ranges of the values of the operating parameters of each category are preset. A neural network is created for each of the divided regions, and the number of nodes of the hidden layer of each neural network is preset, Training data obtained in advance by actual measurement for the values of the operating parameters of a plurality of categories is stored in the storage unit, When the values of the operating parameters of the plurality of categories related to the internal combustion engine newly acquired by the parameter value acquisition unit are within a preset range, the newly acquired values of the operating parameters of the plurality of categories related to the internal combustion engine are used The training data obtained by the actual measurement is used by the computing unit to make output values that vary according to the values of the plurality of types of operating parameters related to the internal combustion engine and values corresponding to the plurality of types of operating parameters related to the internal combustion engine. to learn the weights of the neural network of the divided regions to which the newly acquired values of the operating parameters of the plurality of categories related to the above-mentioned internal combustion engine belong, so that the difference between the training data becomes smaller, When the value of the operation parameter of at least one type among the plurality of types of operation parameters related to the internal combustion engine newly acquired by the parameter value acquisition unit is outside the preset range, the operation parameter of the at least one type of operation parameter is set. The value belongs to a new area defined by the combination of the preset ranges of the values of the operating parameters of each category, and a new neural network is created for the new area, using the newly acquired data related to the internal combustion engine. The training data obtained by the actual measurement of the values of the operating parameters of the individual categories is used to make the output values that vary according to the values of the operating parameters of the plurality of categories related to the internal combustion engine and the output values corresponding to the operating parameters related to the internal combustion engine. The weights of the new neural network are learned so that the difference between the training data of the values of the operating parameters of the respective categories becomes smaller, and each neural network for which the weights have been learned is used to output output values corresponding to the values of the operating parameters of the internal combustion engine. 6. The machine learning device for an internal combustion engine according to claim 5, The operating parameters related to the internal combustion engine are composed of two types of operating parameters, and the value of one of the operating parameters related to the internal combustion engine newly acquired by the parameter value acquiring unit is outside a preset range and is another. When the value of the operating parameter of one category is within a preset range, the new region is outside the preset range for the value of the operating parameter of the one category and belongs to the value of the operating parameter of the other category Within the same range as the above-mentioned divided area of , it is set adjacent to the divided area to which the value of the operating parameter of the other category belongs. 7. The machine learning device for an internal combustion engine according to claim 5, The one before the output layer of the neural network is based on the divided area in the divided area adjacent to the new area excluding the divided area where the number of nodes in the hidden layer before the output layer of the neural network is not set. The average value of the number of nodes in the hidden layer is used to set the number of nodes in the previous hidden layer of the output layer of the new neural network. 8. The machine learning device for an internal combustion engine according to claim 7, The number of nodes in the previous hidden layer of the output layer of the new neural network in the new region is increased in accordance with the increase in the number of input layers where the values of the operating parameters in the new region are input. . 9. A machine learning device for an air conditioner or a secondary battery, comprising an electronic control unit, The electronic control unit has: a parameter value acquisition unit for acquiring values of a plurality of types of operating parameters related to the air conditioner or the secondary battery; an operation part, using a plurality of neural networks including an input layer, a hidden layer and an output layer to perform operations; and storage department, The values of the operating parameters of the plurality of categories related to the air conditioner or the secondary battery are input to the input layer, and the output values that vary according to the values of the operating parameters of the plurality of categories related to the air conditioner or the secondary battery are extracted from the corresponding output layer output, In this machine learning device, For each of the plurality of types of operating parameters related to the above-mentioned air conditioner or secondary battery, a range of values of the operating parameters of each type is preset, and the preset range of the values of the operating parameters of each type is divided into and a plurality of divided regions demarcated by the combination of the divided ranges of the values of the operating parameters of each category are set in advance, A neural network is created for each of the divided regions, and the number of nodes of the hidden layer of each neural network is preset, Training data obtained in advance by actual measurement for the values of the operating parameters of a plurality of categories is stored in the storage unit, When the values of the operating parameters of the plurality of types related to the air conditioner or the secondary battery newly acquired by the parameter value acquisition unit are within a preset range, the newly acquired values of the operating parameters related to the air conditioner or the secondary battery are used. The training data obtained by the actual measurement of the values of the operating parameters of the plurality of categories is used to make the output values changed according to the values of the operating parameters of the plurality of categories related to the above-mentioned air conditioner or the secondary battery to correspond to the The division to which the newly acquired values of the operating parameters of the plurality of categories related to the air conditioner or the secondary battery belong to is learned so that the difference between the training data of the values of the operating parameters of the plurality of categories related to the air conditioner or the secondary battery becomes smaller. the weights of the neural network in the region, When the value of the operation parameter of at least one of the plurality of types of operation parameters related to the air conditioner or the secondary battery newly acquired by the parameter value acquisition unit is outside the preset range, the at least one type is set The value of the operating parameter belongs to a new area defined by the combination of the preset ranges of the operating parameter values of each category, and a new neural network is created for the new area, using the newly acquired and the above The training data obtained by actual measurement of the values of the operating parameters of the plurality of categories related to the air conditioner or the secondary battery is used to change the values of the operating parameters of the plurality of categories related to the air conditioner or the secondary battery by the computing unit. The weight of the new neural network is learned so that the difference between the output value of the air conditioner and the training data corresponding to the values of the operating parameters of the plurality of categories related to the air conditioner or the secondary battery becomes smaller, and each neural network that has learned the weight is used to An output value corresponding to the value of the operating parameter of the air conditioner or the secondary battery is output.

Images